---

^a Environmental Research Institute of Michigan.

^b Calspan corporation.

^c U.S Naval Research Laboratory.

^d NASA Lyndon S. Johnson space Center.

^* Principal investigator.

---

[257] THE SKYLAB EARTH RESOURCES Experiment Package (EREP) sensors provided a multisensor data base for assessing the use of standard information extraction techniques as well as for developing new techniques. The EREP investigators analyzed data from the photographic, electro-optical mechanical, and microwave sensors in ways that were appropriate to a particular application. In several instances, the investigators developed improvements in remote-sensing analysis techniques or provided results that guided the selection of location of operating bands for future sensors.

Because the space-flight performance of the EREP sensors dictates to a large degree the data-processing and analytical techniques used by individual investigators, the in-flight performance of the EREP sensors and the data analysis techniques used for each sensor are summarized in this section. Other techniques are discussed in sections 2 to 5 and in appendix D.

Of all the sensors onboard, cameras filtered for selected wavelength bands produced photographs having the best spatial resolution. The analytical techniques for the EREP photographic experiments reported in this section are discussed in four general categories: visual analysis of the imagery from different spectral bands and from different sensors; microdensitometry and color encoding; multiband image enhancement and analysis, including image digitization and computer techniques; and data processing that provided input parameters for Earth resources management models.

The EREP optical-mechanical scanner provided multichannel computer-compatible tapes that permitted machine enhancement of surface features based on spectral signature analyses. The techniques used to ensure acceptable recognition accuracies included multiband comparisons, ratio processing, spectral classifications based on statistical tests, mixture processing, and signature-extension schemes. An infrared spectrometer provided high-spectral-resolution data in the visible and infrared bands from preselected test sites so that atmospheric corrections and surface signatures could be developed.

Active and passive microwave sensors, operating at the longest wavelengths used by EREP sensors, provided computer-compatible-tape (CCT) data from which information about the atmosphere, the surface roughness, the soil moisture, and the shape of the Earth could be determined. To achieve the full performance potential of the altimeter, special techniques were developed to determine the antenna pointing angle directly from the radar return and to calibrate range and power output for pulsewidth- and beamwidth-limited conditions. The interpretation of data from the radiometer/scatterometer (RADSCAT) required different statistical methods of analysis for verification and correction. Rigorous methods for computing ocean emission at microwave frequencies for different physical conditions were developed and used to separate various physical parameters from the radiometer data characteristics.

[258] The data analysis techniques provided for the conversion of sensor data into meaningful application information such as numbers, thematic maps, and boundaries from which decisions could be made, action taken, and new programs initiated. In general, this transformation could not be accomplished without supporting information, surface truth, or a framework of model analysis. The development of suitable models and frameworks for inputting this new form of data supplied by spaceborne sensors is one of the main challenges to the maximum utilization of remote-sensing techniques.

SENSOR FLIGHT PERFORMANCE

The EREP investigators defined their experiments on the assumption that predicted sensor performance levels would be achieved during flight. Therefore, an important task during the EREP data passes was to test the sensors to determine their actual operational performance.

Sensor performance was evaluated in three different areas: (1) functional performance (e.g., camera shutter operation and antenna movement), (2) geometric performance (e.g., spatial distortions in the imagery, pointing accuracy of the antenna), and (3) radiometric performance (e.g., radiometric accuracy and precision). Generally, only the flight data collected for the EREP Principal investigators were used for the sensor performance evaluation; however, for the lunar or deepspace calibration studies, a restricted amount of data was collected specifically for sensor performance evaluation. Detailed discussion of the performance and engineering evaluation of each sensor is published in an NASA internal document ¹ and summarized by Potter et al. (ref. 6-1). A brief summary of the results of the sensor evaluation is presented for each sensor (app. A).

Multispectral Photographic Camera

The Multispectral Photographic Camera (S19OA) consisted of six boresighted camera stations, four with spectral filters and black-and-white film, one with high....

---

....resolution color film, and one with color-infrared film. An example of imagery from the color-infrared camera station showing the imperial Valley and the Salton Sea, California, is illustrated in figure 6-1. This figure is an enlargement; one entire S190A frame covers a 163-km-square area. Vegetation is delineated by red tones. The same scene, photographed with the high-resolution color camera station, is shown in figure 6-2. To demonstrate the resolution capability of the S19OA system, this scene was enlarged. The limit of resolution can be seen in figure 6-3. The measured resolution capability of the camera for this film was approximately 27 m per line pair for high-contrast sites and approximated the value expected from preflight tests. The registration capability of the black-and-white imagery is illustrated in figure 6-4, a composite generated by superimposing enlarged images from three black-and-white camera stations for the imperial Valley scene (figs. 6-1 and 6-2). The images can be registered to within better than 20 m on the ground, approximately the value expected from preflight tests. The S190A camera was a radiometric....

....camera in that film density could be related quantitatively to the intensity of radiation incident on the lens. In-flight measurements using the ground sites and the Moon established radiometric calibrations accurate to within ± 30 percent.

Earth Terrain Camera

The Earth Terrain Camera (S19OB), a single camera station, was designed to demonstrate the application of high-spatial-resolution imagery in Earth resource surveys and normally was operated with either high-resolution color or black-and-white film. An S190B image made with high-resolution color film covering a 109-km-square area over Phoenix, Arizona, is shown in figure 6-5. The Sun City Development area, which is used to demonstrate the resolution capability of the camera, is visible in the upper left center of the photograph. Figures 6-6 to 6-8 show successively greater enlargements of this same area. The resolution limit of the camera system was approached at the magnification shown in figure 6-8. The measured resolution limits were approximately 9 m per line pair from black-and-white film. Performance of the camera was within expected limits for all parameters.

Infrared Spectrometer

The Infrared Spectrometer (S191) measured the radiation from a 0.46-km-diameter area on the Earth in the ranges 0.4 to 2.5 µm and 6.6 to 16 µm. The telescope which collected and transmitted light to the spectrometer could be pointed by a crewman at preselected sites below the spacecraft. Photographs of the scene through the telescope recorded pointing angle and time so that the position of a measured site on the ground could be computed.

The S191 was tested by measurements of calibrated ground sites and areas of the Moon. Typical ground site spectra taken with the spectrometer system are shown in figure 6-9.

Careful analysis of the spectral data indicated that the instrument performance was within preflight specifications in the reflective range, except at the blue end of the spectrum (0.4 µm), where excessive off-band radiation appeared. When viewing similar scenes, relative values in the thermal emissive part of the spectrum were within expected tolerances, but the absolute radiance values were incorrect by variable amounts equivalent to a minimum of 1 K to several kelvins. Postflight analysis of data from the backup spectrometer flown on a helicopter indicated that the probable cause of this error was an incorrect transmission coefficient for the dichroic beamsplitter, which divided the reflective and emissive parts of the spectrum.

Multispectral Scanner

The Multispectral Scanner (S192) provided imagery in 13 spectral bands ranging from the blue (0.4 µm) to the thermal-infrared (12.5 µm) portions of the electromagnetic spectrum. The image data were recorded on magnetic tape for later analysis by high-speed computers. An example of data from the S192 is shown in figure 6-10. Three of the 13 bands (2 infrared and I green) were superimposed to make this image. Agricultural features, such as circular irrigated fields,....

---

---

....can be seen clearly in the imagery. Close examination of figure 6-10 reveals that the unfiltered imagery contains a noticeable degree of noise. To determine the effects of this noise on the computer analysis of this imagery, the small outlined area was subjected to computer analysis. The results of this analysis are shown in figure 6-11. The analyzed area is outlined on an S190 photograph on the right, and the results of the computer classification are shown on the left. Agricultural and other features were classified with an acceptable accuracy (greater than 90 percent. The noise in the imagery did not have a major effect on computer classification; apparently, some of the noise was correlated between the different spectral channels. However, the noise was larger than had been expected on the basis of preflight measurements. Consequently, a noise analysis was performed, and two major types of noise were found: high frequency ( 20 kHz) and low frequency ( 20 Hz). in several channels, high-frequency noise at discrete frequencies....

---

....was observed. This noise produced herringbone patterns in the imagery. Because the noise occurred at sharp, well-defined frequencies, it was possible to remove the noise by computer analysis (mathematical filters).

The low-frequency noise was most noticeable in the thermal-infrared channel and produced heavy banding (random noise) in the image. This imagery was not useful for photointerpretation because the banding obscured surface features. Consequently, mathematical filters and every-scan calibration data were used to minimize or remove the noise.

An example of thermal imagery is illustrated in figure 6-12. Low-frequency noise was present before data processing; after processing, the low-frequency noise had been filtered out. improvement in the quality of imagery for photointerpretation is evident.

The thermal imagery acquired on the first two missions was acceptable for some applications but not for others. As a result, the detectors flown on the Skylab 2 and 3 missions were replaced with an improved thermal detector by the Skylab 4 crewmen. Thermal imagery from the Yuma, Arizona, region (fig. 6-13) was obtained with the improved thermal detector. A substantial improvement is evident in the image quality produced by the detector. The radiometric performance of this thermal detector was such that temperature differences of approximately 0.8 K were equivalent to detector noise.

During initial in-fight checkout and operation, difficulties with the manual alinement and focusing of radiation on the detectors were encountered. In addition, some of the bands had an excessively small dynamic range. Therefore, contingency attenuators.....

....were installed by the crewmen. A small percentage of data was lost because of these problems. All other performance factors of the system (band-to-band registration, geometric resolution, radiometric calibration, etc.) were within the limits expected from the preflight calibrations.

Microwave Radiometer/Scatterometer and Altimeter

The Microwave Radiometer/Scatterometer and Altimeter (S193) operated at the K_E-band (2.2 cm) wavelength in three different modes: as a passive radiometer, measuring the microwave energy emitted by the Earth; as an active scatterometer, which sent microwave pulses to the Earth and measured the intensity of the return echo; and as a radar altimeter. All these functions shared the same antenna, which was gimbaled to permit scanning of the Earth in various patterns. The antenna functioned acceptably during the Skylab 2 mission and most of the Skylab 3 mission; however, near the end of the Skylab 3 mission, the antenna began scanning erratically. Early in the Skylab 4 mission, the crewmen were able to restore the antenna scan in one direction by mechanically pinning it in one axis and by making modifications to the electronic circuits. However, Skylab 4 data proved to be anomalous and very limited amounts of the data were recoverable for the investigators.

Analysis of radiometer performance indicated that emitted radiant power was measured with an accuracy of at least 4 percent (corresponding to a brightness-temperature accuracy of ±7 K) and with a precision of at least 2 percent (or ±1.5 K) for typical ground scenes. The scatterometer performance showed that the reflected signal was measured within an accuracy of 4 percent for typical ground scenes, with a precision of at least 2 percent. With this accuracy and precision, the scatterometer was capable of measuring reflected signals that varied in amplitude by a factor of greater than 10 000:1.

Altimeter performance was measured from analysis of the data obtained over regions with known surface contours. The accuracy of the altitude measurement was ±7 m with a precision of ± 1 m.

With the exception of the antenna, all performance aspects of the S193 during the Skylab 2 and 3 missions equaled or exceeded predicted values.

---

[267] L-Band Radiometer

The longest wavelength (21 cm) sensor onboard the Skylab vehicle was the L-Band Radiometer (S194), which measured radiant energy from an approximately 124-km-diameter area of the Earth. Effective measurements were limited to a 111-km-wide swath centered about the nadir point. The performance factors of most interest for this passive microwave sensor were its antenna pattern, pointing accuracy, and radiometric accuracy.

Comparison of predicted and observed signals from well-defined surface features were used to define the S194 antenna pattern and pointing accuracy. Both these factors were within a few percent of predicted values. Radiometric performance was determined from measurements of the Sahara Desert, where large sandy areas that have uniform emissivity and temperature exist. A precision of approximately ±0.2 K and an accuracy of approximately ±1.7 K at 285 K were found.

THE S19OA AND S19OB PHOTOGRAPHIC IMAGING SENSORS

The Skylab Program provided a unique opportunity to Use photographic imagery in Earth resources investigations. Investigators were able to analyze high-resolution photographs processed under carefully controlled conditions, with known camera performance characteristics, and with a wide range of film and spectral filter combinations. The investigators used both sophisticated data analysis and standard photointerpretation techniques. The advanced techniques included comparison of image information among spectral bands, precise image densitometry, and multiband analysis and enhancement techniques. The application of these sophisticated analysis techniques led to significant results that are discussed in the preceding discipline summary sections.

Almost all the analysis techniques and results are based on two key properties of the Skylab photographs. A brief description of these properties will assist in the understanding and appreciation of most of the techniques developed during the photographic experiments.

First, the Skylab photographs provided investigators with information in many wavelength, or spectral, regions. The spectral data were obtained from filtered black-and-white films or from information in the layers of multiband films, such as color and color-infrared film. In the case of filtered imagery, the spectral bands were defined by the combination of the wavelength transmitted by the filter (filter spectral bandpass) and the wavelength sensitivity of the film (film spectral sensitivity). For the multiband films, the spectral bands were determined primarily by the spectral sensitivity of the individual film layers. The investigators extracted the spectral information from the multiband imagery by separating the information in the individual film bands through a filtering process. Thus, by using either the filtered or the multiband imagery, the Skylab investigators had a wide variety of high-resolution photographic images for Earth resources investigations.

Second, the Skylab photographic experiment package was designed to enable use of the cameras to measure the amount of energy reflected from the Earth. In other words, the Skylab cameras could be used as imaging photometers (devices that measure incident energy). Thus, investigators were able to determine the energy coming from any part of the photographic scene by careful measurements of the density of that part of the image on the exposed film.

Each roll of film had a sensitometric control strip (fig. 6-14) that could be used by the investigator to relate photographic density to the energy that exposed the film. Each control strip consisted of a set of film areas exposed with known amounts of energy. A comparison of the density from a Skylab scene with the same density on the control strip provided a measurement of the energy incident at the film plane in the spacecraft. Also, because the effects of the camera lenses and shutters on exposure energy had been carefully measured, camera effects could be removed from the image data. The photographic image data could be processed to obtain an accurate measure of the energy reaching the spacecraft from any portion of the photographic scene. Thus, the investigators not only had high-resolution photographs that enhanced terrain and resource features, they were also able to extract a measurement of reflected energy in each spectral band being studied and to compare the energy differences among the spectral bands. The ability to measure energy values in the different spectral bands was of.....

....fundamental importance to several new techniques developed during the photographic experiment.

The data analysis techniques developed from the S190 photographic experiment can be grouped into four categories: (1) techniques arising from the application of conventional photointerpretation methods to the unique data format provided by the Skylab photographs, (2) techniques evolving from various forms of image densitometry, (3) techniques developed from multiband spectral enhancements or analyses, and (4) techniques in which Skylab photographs provided a unique model input for Earth resources analysis.

Photointerpretation Techniques

Conventional photointerpretation techniques principally involve visual evaluation of shapes and patterns on the photographic image (appendix D). The shapes and patterns are defined by the tonal and textural variations within the scene. Interpreters evaluate the patterns discriminated on the basis of their understanding of, and experience with, the physical processes under study. Many successful applications of such conventional photointerpretation analysis can be found in the individual discipline sections of this report. The new approach for Skylab visual analysis techniques is basically an extension of the conventional photointerpretation methods to include, or take advantage of the unique characteristics of Skylab photographs; namely, the perspective achieved from space (large areal view with high resolution), a wide selection of spectrally filtered images, and the photographic format from which visual comparisons can be made.

Each frame of Skylab imagery covers a large area; the S190A photographs cover approximately 160 km on each side, and the S190B photographs cover approximately 110 km on each side. A high-altitude aircraft flying at 15.24 km with a standard 150-mm focal length, 230-mm format mapping camera would require 50 photographs to cover fully an area 160 km on each side, assuming perfect flight lines and no overlap between photographs. When a standard amount of photographic overlap is allowed, the number of photographs required to cover the 160-km square increases to more than 200. A prodigious effort would be required to correlate these photographs so that an interpreter could recognize large-scale effects occurring over 16 to 160 km.

The Skylab photographic experiment removed this logistical obstacle by providing high-quality, high-resolution photographs covering large areas. Photointerpreters could thus evaluate phenomena at physical scales previously unattainable. For example, figure 6-15 is an S190B color photograph of the Pacific Ocean near Point Arena in northern California. The whitish tones in the water are caused by sediment of varying concentrations. A study of this photograph by Pirie and Steller (ref. 6-2) provided information on the occurrence of upwellings and location of coastal current that would have been impossible from aircraft photographs.

Mapping from aircraft photographs would have required a degree of flight-line accuracy that would have been extremely difficult to achieve over the ocean. The ocean has no landmarks with which to sight a flight line or to orient photographs. Also, aircraft coverage would have required hours of flying during which the currents could change or lighting conditions could vary. The variation in lighting conditions would cause analysis difficulties because current structures were extracted on.....

[270] ....the basis of tonal variation over the ocean. Finally, each aircraft photograph would have brightness changes from side to side across the format caused by differences in angles of illumination and reflection. These brightness changes would also have contributed to analysis difficulties. The problems of flight-line accuracy, time variation, and tonal or brightness fluctuations were minimized by Skylab photographs.

Barnes et al. (ref. 6-3) evaluated the usefulness of EREP data for mapping snow cover. Accurate snowcover mapping is crucial for runoff prediction and water management. These investigators determined that S190 color and color-infrared photographs could be used to map snowpack extent more accurately than could images provided by any other spacecraft or aircraft system. A key element in the S190 mapping capability was the perspective attributable to large-area coverage. Figure 6-16 depicts snowline regions in central Arizona determined from an S19OB photograph and from a concurrent aerial snow survey. The aerial survey snowline is considerably less detailed than the Skylab-delineated snowline, and the position determined from the aerial survey does not fit the topography as closely as the Skylab snowline. The accuracy of runoff prediction is closely related to the accuracy of snowline mapping.

Image analysis techniques using photointerpretation of image patterns were significantly extended in the Skylab photographic experiment by including comparisons of photographic image patterns with patterns found on nonphotographic imagery, such as those from the S192 Multispectral Scanner. Skylab investigators were thus able to extract additional information on such problems as the detection of oceanic upwelling regions for fishery resources, the determination of vertical sediment distribution in coastal waters, the detection of melting snow regions for runoff prediction, and the differentiation of water droplet clouds from snow.

Szekielda (ref. 6-4), for example, used S190 color photographs to study ocean color changes as possible clues to the location of an upwelling. As discussed in section 5, upwelled waters afford an excellent medium for the growth of fish. Monitoring the location of upwelling regions from space is thus extremely important.

Photointerpretation of red and green spectral bands proved useful in determining the vertical distribution of suspended sediments (ref. 6-2). Because penetration of water by red wavelengths is small, an image in the red spectral band depicts only near-surface information. Water penetration in the green spectral region is greater than in the red, and an image in the green spectral region provides detailed information on subsurface sediments. The discrimination of such surface and subsurface effects is depicted in figure 6-17.

Barnes et al. (ref. 6-3) developed a technique for the automatic discrimination between snow and water droplet clouds by comparing S192 imagery in the visible spectral region to that in the middle-infrared (1.6 to 2.4 µm) region. In the latter spectral range, the reflectance of snow is almost zero, whereas water droplet clouds have a high reflectance. A comparison of the visible spectral region, in which both types of objects have high reflectance, with the 1.6- to 2.4-µm region thus permits discrimination. The discrimination technique is of particular significance for automatic snow mapping because cloud discrimination has been recognized as a serious hindrance to eventual machine processing of satellite data for snow-cover mapping.

Similarly, Piech et al. (ref. 6-5) demonstrated that a comparison of red (0.6 to 0.7 µm) and near-infrared (0.8 to 1.0 µm) spectral bands was useful in differentiating atmospheric turbidity that occurs over a large lake from the turbidity that occurs within the lake. The technique is particularly useful in accounting for the effects of very light, wispy clouds or haze. In the near-infrared spectral region, a lake without high sediment load will appear black, whereas a cloud or haze effect will maintain a signal or pattern similar to its pattern in the red spectral band. A comparison of the two spectral bands thus permits discrimination between atmospheric effects and lake effects; this discrimination is important in measuring parameters such as turbidity and chlorophyll concentration.

A comparison of visible and near-infrared bands has also proved useful for detecting snow cover in a melting condition (fig. 6-18). Such snowmelt information is particularly needed in regions where significant portions of the snowpack can melt within a few days, or even hours. A melting snowpack decreases in apparent extent when viewed in the visible and near-infrared bands because of a decrease in the reflectance of melting snow. A comparison of visible and near-infrared bands yields a measure of the melting condition of the snowpack.

Densitometric Analysis Techniques

A photointerpreter is limited in the ability to discriminate tonal differences within an image. The limitation arises because of visual and psychological constraints in quantifying contrast differences between points of an image that are separated or between points of an image Iying in an area of changing image texture. Several Skylab investigators found it desirable to quantify important density or exposure changes occurring within the photographic scene by using various forms of image densitometry.

Some investigators (ref. 6-5) used macrodensitometers to measure density values over a large spot on a photograph, typically approximately 0.5 mm in diameter. A density reading on a O.5-mm-diameter spot covered a ground area of 1.5 km in diameter on S190A photographs and 0.5 km in diameter on S19OB photographs. Other investigators (refs. 6-2 and 6-5) used microdensitometers to measure densities over spot [273] diameters of approximately 25 µm corresponding to a ground dimension of 75 m in diameter on S19OA photographs and 25 m in diameter on S19OB photographs. The microdensitometers were used in a manner similar to the use of a macrodensitometer, to measure the density of specific image spots, or to scan the image to provide a trace, or record, of density values from which contours or shapes could be drawn.

Another method used for image densitometry was the application of color-encoding devices, in which a television vidicon scans the photograph and converts the light signal to density values. The density range of the scene is then divided electronically into 10 or 20 ranges of density. Each density range is assigned an individual color, and the photographic scene is redisplayed on a color television screen with the photographic densities encoded as colors. The resulting color patterns can then be interpreted as representing different types of trees or differing sediment concentrations, depending on the problem under investigation.

Densitometry of individual layers of the multiband color or color-infrared films was accomplished by two methods. In one method, a filter transmitting light from one film layer was inserted into the densitometer so that density values were measured for only that film band or layer. In the second method, a black-and-white copy, or separation, of the information on one film layer was made through a filter. Densitometry was then performed on the black-and-white copy of the scene. Regardless of the techniques used, investigators were able to obtain spectral-band information from the multiband films and to apply these data to their resource investigations successfully.

The densitometric analysis techniques were applied not only to Skylab photographs, but to various enhancements of the photographs such as ratio combinations of film layers (ref. 6-5). Application of densitometry to such multiband enhancements is discussed in section 3 and in the following subsection. Densitometry of the original photographs proved most useful in water quality studies related to coastal dynamics and sediment transport, and in land use studies such as timber differentiation and classification.

Baldridge et al. (ref. 6-6) applied color encoding of color and color-infrared imagery to timber differentiation and analysis. Color encoding was successfully applied to timber mapping, differentiation of hardwood and softwood stands, and evaluation of timber maturity. The infrared-sensitive layer of color-infrared film proved most sensitive and valuable for timber differentiation. Figure 6-19 contains examples of successful delineation of timber maturity from S19OB color-infrared photographs. This figure shows color encoding of a sample area in Mahoning County, Ohio, containing only three stages of stand maturity. The area was 1 of 16 sample areas in which machine determination of stand maturity from S19OB photographs was undertaken. Forest management personnel confirmed the validity of the classification based on the color-encoding techniques in all the sample sites.

Numerous investigators found densitometric analyses of significant value in studies of coastal and estuarine dynamics and of sediment transport (refs. 6-2, 64, and 6-7 to 6-9). Welby and Lammi (ref. 6-9) used color-encoding techniques to interpret underwater and shore topography. The Skylab photographs also proved useful in revealing sediment in water bodies. The effect of a period of precipitation on sediment discharged into a major lake was recognized, and the use of orbital photographs for measuring relations between rainfall and sediment load in a drainage basin was shown to be possible.

Szekielda (ref. 6-4) related color changes to particulate concentrations in the study of upwelling areas off the coast of northwest Africa. Gordon and Nichols (ref. 6-8) successfully used color encoding and microdensitometry to analyze southern Chesapeake Bay water color and circulation. Red- and green-band data were found to be the most useful for mapping water types related to transparency, turbidity, and suspended-sediment load. It was further established that suspended sediment could be used as a tracer of water movement to discriminate small-scale mixing patterns and local tidal currents. Large-scale patterns were shown to reflect bottom topography indirectly because the bottom serves as a source of suspended material.

A near-linear relationship between reflected radiance and suspended solids in the concentration range of 20 to 80 p/m was found by Yarger and McCauley (ref. 6-7). Pirie and Steller (ref. 6-2) found that variations in sediment concentration could be observed. Figure 6-20(a) is an S19OB photograph of the San Francisco Bay area during a period of high sediment discharge. Microdensitometer traces in the green spectral band are shown in figure 6-20(b), with contour intervals ranging from highest reflectance and highest suspended solids (1) to lowest reflectance and lowest suspended solids (4). Before the Skylab overpass, an iridium tracer was added....

....to the sediment flow. When fresh and saline waters meet (as in San Pablo Bay), flocculation occurs. Measurement of the amount of tagged sediment in subsequent dredging should therefore correspond with the sediment flow pattern and variations. Figure 6-20(c) indicates the correlation between the brightest regions of flow and the amount of tagged material observed in subsequent dredging of San Pablo Bay. The correlation between the densitometer map and dredging results is very good.

Analyses Using Multiband Spectral Enhancements

Several investigators used data processed from two or more spectral bands. The individual-layer data were combined, either photographically or electronically, with data from other photographic bands to generate displays for interpretation. Hardy et al. (ref. 6-10) investigated land use mapping through combinations of three S190 spectral bands processed on diazo photographic materials to produce maximum color contrast among different land use categories. Colwell et al. (ref. 6-11) digitized densitometric data from four S190 bands and subsequently performed a discriminant analysis on the multiband data to classify crops successfully. Piech et al. (ref. 6-5) used image microdensitometry to determine atmospheric and processing corrections, then photographically and electronically divided two spectral bands (which were corrected for atmospheric and processing effects) to measure eutrophication indices of...

....large lakes. The multiband analyses can therefore be grouped into photographic data reduction, electronic data processing, and a combination of electronic and photographic data analysis.

Photographic data processing.-Hardy et al. (ref. 6-10) used S190 photographs to update an existing New York State land use inventory. A key aspect of this investigation was the development of an inexpensive, systematic procedure for generating enhanced color composites from the filtered S190 photographs. A color prediction model was developed to automate the selection and generation of color composites that maximized color contrast between selected land use categories.

Enlargements were made from the Skylab photographs, and the density range and contrast of the enlargements of each spectral band were standardized, or normalized, so that the density of terrain features would correspond to the exposure range of diazo photographic materials. The diazo photographic materials, which are usually used for line-drawing reproduction, were selected because they are inexpensive to purchase and process.

Density values obtained from each spectral band for known areas representing different land uses are fed into the computerized color prediction model. The model relates the density values in each spectral band for each category of land use to the spectral properties of the various diazo films. The model then produces a combination of spectral-band data and exposure data to maximize the color contrast among the land use catego-....

....-ries being examined. Three land use categories can be examined for each composite; a set of color composites used for land use interpretation is shown in figure 6-21. Use of the color-composite data resulted in aggregate errors of approximately 12 percent for Level I land use classification and 25 percent for Level II classification.

Electronic data processing.-Automatic processing of Skylab photographs for crop classification was evaluated by Colwell et al. (ref. 6-11) and Silva (ref. 6-12). in the land use studies, computer processing of photo. graphic density data that had been converted to digital form by scanning with a microdensitometer was used.

The four black-and-white bands of the S19OA camera (green, red, and two near-infrared spectral bands) were scanned with a microdensitometer using a spot size approximately equivalent to 0.41 hm² on the ground (Colwell et al., ref. 6-11). The density measurements were recorded on magnetic tape and processed using pattern recognition algorithms.

The optimum combination of spectral bands was determined, as were interclass divergences, from combinations of crops. A classification map based on the training statistics was then generated. Finally, the data were reprocessed to produce a map wherein a "nearest neighbor algorithm" was used and accuracies of classification were computed.

The classification algorithms were applied with surprising success to data from a set of vegetable crops in the Salinas Valley, California. Analysis of the spectral densities in the four bands permitted classification to an overall accuracy of 49 percent early in the crop cycles and 85 percent later in the crop cycles. The latter accuracy is excellent considering that multidate information was not used in the crop classification.

Silva (ref. 6-12) also digitized S19OA photographic data using both color-infrared and black-and-white....

[281] ....multiband frames. The scanning spot was 25 µm in diameter, and lines were scanned at 20-µm intervals. A Kodak 92 filter was used to separate the red layer, a Kodak 93 filter for the green layer, and a Kodak 94 filter for the blue layer. The equivalent ground size for the aperture used was approximately 68 m in diameter. The four black-and-white films of the S190A were also digitized but without filters. A total of 256 different levels of film density was measured, and level 255 was set for the darkest area between frames. The three-channel digitized set defined by the three dye layers of the color-infrared transparency maintained good spatial registration because it was produced from the same transparency. The four-channel digitized set from the four black-and-white films of the S190A camera had to be registered by subjecting an 1100- by 1100-line block of data to a second-order least squares ft using 60 points evenly distributed across the area. The registration was 40 and 70 m for the pair of visible and infrared channels, respectively, but the registration is within two picture elements (pixels) between the visible and infrared bands because of differences in resolution between the two types of emulsion.

Once in proper format, the data were clustered and training fields for each of 12 classes were selected. The digitized color-infrared data were superior to the digitized black-and-white data and had classification accuracies of approximately 80 percent. Some degradation in results was caused by the shoreline of a lake and the narrow course of a river, which did not provide an adequate training set and would be affected by the misregistration between channels.

The successful application of automatic classification from multiband photographic imagery is important because of two characteristics of the photographic data: (1) ease of data storage, which permitted the full multispectral scene from the S190A camera system to be recorded and stored on four 5.72-cm black-and-white film transparencies, and (2) the excellent geometric fidelity of the photographic imagery, which facilitated registration and use of multidate imagery.

Photographic and electronic data processing.-Piech et al. (ref. 6-5) used microdensitometry of image elements to measure atmospheric effects and subsequently to reduce and process the multiband data in the form of target spectral reflectance values. The multiband reflectance analyses were used to obtain the eutrophication indices of large lakes such as Conesus Lake, New York; the relative value, or ratio, of various reflectance bands is related to eutrophication indices of the lakes (fig. 6-22). The reflectance of a lake is typically approximately 2 to 3 percent, whereas the atmospheric component of the signal at the Skylab spacecraft can have a reflectance of approximately 10 percent. A small change in atmospheric properties from one sampling date to another could thus be misinterpreted as a significant change in lake properties. The accurate measurement of relative reflectance values between spectral bands therefore requires accurate removal of atmospheric signal noise in both spectral bands.

The results of the investigation demonstrated that image microdensitometry could be used to specify accurately the atmospheric component of exposure. Figure 6-23 depicts the relative values of blue. and green-band lake reflectance as measured from the Skylab spacecraft compared to those measured from an aircraft underflight at an altitude of 3048 m. The aircraft measurement techniques had previously been shown to agree consistently with ground measurements of spectral reflectance ratios. The Skylab and aircraft data agree within the system measurement accuracy. The investigators pointed out that increased accuracy in the measurement of reflectance values would be obtained by an increase in image resolution.

Subsequent reduction of the lake spectral data was accomplished by means of a combination of photographic and electronic data processing. A black-and-white photographic copy, or separation, of each spectral band of the S190A color photograph was modified according to the atmospheric effect measured for that spectral band; am appropriate increase in image contrast in each spectral band removed the effects of the atmospheric signal for that spectral band.

---

A reversal copy was then made of one of the corrected spectral bands for which relative reflectance values were desired. An overlay of the positive copy of the other spectral band together with the reversal copy comprised a density map proportional to the reflectance ratio values of the lake scene. The data were then extracted, or displayed, using a color-encoding device. The crucial elements to the multiband analysis, however, were the measurement and the removal of atmospheric effects.

Approaches and Model Inputs

The Skylab photographic experiment served as a focal point for several investigations by providing a medium through which large areas could be studied by multidisciplinary teams and through which successful analysis methods could be verified and extended. Although the spectral properties of the Skylab photographs were used in all these studies, the most important characteristics were large-area coverage and high spatial resolution.

Skylab photographs enabled performance of an interdisciplinary study of the hydrology of prehistoric farming systems within a large and environmentally diverse area of central Arizona (Gumerman et al., ref. 6-13; fig. 6-24). Hydrologists, geologists, biologists, and archeologists evaluated the adaptation of prehistoric humans to the semiarid desert of central Arizona, and their creation of land management and water control systems. Such an analysis required data over thousands of square kilometers and an understanding of the numerous variations in environment over this region. Coverage of such a large region using only ground activities is impossible. Data from Skylab or high-altitude-aircraft systems provide the archeologist with a significant regional perspective for understanding surface geology, hydrology, and vegetation patterns and enable formulation of specific questions and relationships among the environment, prehistoric settlements, and subsistence systems.

The study covered an area extending from the Lower Sonoran Life Zone just north of Phoenix to the Upper Sonoran Life Zone just south of Prescott, including a biological transition zone between the desert floor and the plateau. Ecologically significant subareas, or drainage basins, were selected for study on the basis of basin area, stream length and order, slopes, bedrock type, and rainfall distribution. Estimation of available water was established from these parameters and from vegetation communities. With this information, it was possible to predict with some accuracy the basins most likely to have prehistoric water management systems. Delineation was based primarily on the topography and geometric characteristics of the drainage basins. The drainage information from the Skylab photographs, together with slope and landform data from high-resolution U-2 aircraft imagery, enabled archeologists to discern potential areas of prehistoric utilization.

Within the basins most suited for agricultural purposes, a series of exploitation models for one study area (fig. 6-25) was developed by using the topography, slope, and drainage patterns interpreted from Skylab and U-2 photographs. A decision model developed by Plogg and Garrett (ref. 6-14) was used to evaluate the alternative exploitation schemes for the entire area. The model was informally tested using the remote-sensing interpretations together with statistical data from the traditional sources as inputs. The results provided the archeologists with a fairly comprehensive understanding of the cultural and economic patterns for a large portion of the central Arizona area.

[285] An extension of the precision of multistage sampling of timber volume to a Skylab data base was accomplished by Langley and Van Roessel (ref. 6-15). in multistage sampling, small-scale photographs are used to delineate regions of homogeneous characteristics, large-scale photographs are used to specify field plots of the sample classes for ground measurements, and subsequent ground measurements of yield or volume are used to develop the field plots of the sample classes. The investigators developed a resection technique for the Skylab photographs that permitted accurate location of sample-unit boundaries on the photographs. The subsequent analysis of the inventory precision resulted in the determination of increased economy and efficiency attributable to the space photographs. The precision of the volume estimate could be further improved through the use of multidate photographs; however, Langley emphasized that an increase in resolution to the point at which individual trees can be discriminated would provide the most significant increase in timber-volume accuracy.

THE S192 MULTISPECTRAL SCANNER

Although photographic systems provide better spatial resolution because of simpler construction and the availability of good-quality lenses and modern high-resolution films, optical-mechanical imaging sensors have four advantages relative to the usual photographic system: (1) the sensors operate in spectral regions not available to photographic systems (i.e., generally beyond 1 µm), (2) the position and width of each spectral band can be specified and controlled, (3) the sensors operate simultaneously in several spectral bands wherein each pixel is in both position and time registration across all spectral bands, and (4) the data are better calibrated. As explained previously, the S19OA camera system provided registered multiband imagery with color and color-infrared film, and was radiometrically calibrated. However, the spectral bands overlapped to a degree, and their location and width could not be altered. When the multichannel electrical analogs of the scene are recorded on CCT's, a wide range of signal-processing techniques becomes available. Modern, high-speed computers can process the large volume of data generated by satellite sensors.

Thirteen bands of information between 0.4 and 13 µm recorded in the S192 differed from the four-band Landsat systems that operate between 0.5 and 1 µm. A considerable amount of theoretical work involving statistical decision theory has been applied to the processing of aircraft multispectral scanner data with varying degrees of success. The Skylab S192 sensor was designed to apply these sophisticated schemes to space data, in which both resolution and atmospheric effects were different from those encountered at aircraft altitudes. The question concerning the optimum bands as seen from space for specific disciplines needed to be addressed.

Skylab investigators used S192 data as composite images and computer-compatible tapes to derive parameters of interest for agriculture, water resources, oceanography, land use, and geological and hydrological problems. A few investigators determined the effects of atmospheric attenuation and scattering on the uniqueness of signatures for different classes of objects and on the problems of signature extension.

The unique advantages of the S192 Multispectral Scanner form the basis for discussion of image analysis techniques. In some cases, investigators applied old techniques to new problems, whereas in others, new methods were developed for using the multispectral scanner data. The data analysis techniques are organized into four groups.

1. Single-spectral-band techniques-The investigator used the information in a single spectral band.

 

2. Two-spectral-band techniques-in some studies, the investigator required the use of only two spectral bands.

 

3. Multiple-spectral-band techniques-in these investigations, multiple spectral bands, three or more, are used either to generate false-color imagery or to perform multispectral pattern recognition.

 

4. Imagery use in Earth resource models-For example, a meteorological model may incorporate surface temperature as a key parameter in modeling phenomena such as winds, clouds, or rainfall. The data from the scanner can be used to input surface temperatures into the model and thus infer the meteorological conditions predicted by the model.

Single-Spectral-Band Techniques

The radiometric properties of a single spectral band can be advantageous in some applications research. The general approach for use of single-band radiometric data was to develop empirical correlations between the radiometric intensity and a specific property of the sur-....

...-face. By the careful choice of spectral band, a great variety of surface conditions can be mapped, including soil salinity, land use patterns, and lake outlines. In addition to radiometric intensity data, the single-spectral band imagery contains information on the spatial distribution of intensities. Fourier transformation can be used to analyze the spatial frequencies in single-band image data and to correlate them with geologic or hydrologic features of the surface. In the following paragraphs, specific examples of these techniques are summarized.

Maul et al. (ref. 6-16) used S192 imagery to study cloud features over the Atlantic Ocean between Florida and Cuba. One analytical technique used was the automatic detection of clouds to specify free areas for further sea-surface temperature or ocean color measurements. Band 8 (0.98 to 1.08 µm) was chosen because of its high atmospheric transmissivity and high water absorption. Figure 6-26 shows the frequency distribution for the water and cloud signals in band 8. The cluster around the low radiances represents water, whereas the cluster at the high end represents clouds and land. Previous work (ref. 6-16) established the Gaussian distribution of the radiance N reflectance from the ocean; therefore, by a suitable ft of the data for the water reflection, a radiance range of three standard deviations around the mean radiance = 0.5 µW/(cm^{2 .} sr) should include all cloud-free pixels.

The range is then spread over the available dynamic range of the display by the conditions (for a negative image)

(6-1)

where M is the maximum value allowed by the display, is the mean radiance of the cloud-free data, is the standard deviation, and k is an arbitrary constant. By setting k = 2, 95 percent of the cloud-free data would be recorded over the full range of the display. Detection of anomalies on the ocean surface would be improved by this method of enhancement.

The problem of area measurement using digital values and finite spatial resolution leads to errors for those pixels containing parts of two classes. Gilmer and Work (ref. 6-17) faced this problem in attempting to measure lakes and pond areas and their changes. Usually, all water areas are underestimated. Using the 1.55- to 1.75-µm band of the S192, pixels within the boundary of each lake in their test site were counted. In terms of percentages, higher errors were more frequent for smaller ponds and for those with irregular shapes. Because of the conical scan, the actual errors varied depending on the size and shape of the conical scan direction.

[287] Techniques involving Two Spectral Bands

The intensity of radiation received by the multispectral scanner from a given element of the Earth's surface is influenced primarily by the reflectance of the surface and the Sun angle. However, other factors such as the angle or slope of the surface and the amount of atmospheric haze or cirrus clouds over the surface influence the intensity of received radiation. To a first approximation, the effects of illumination, angle, haze, and clouds can be suppressed by dividing the intensity in one band by the intensity of a different spectral band. The band ratio is largely a function of the difference of surface reflectivity at the two wavelengths. Several investigators used this general technique in their investigations.

Vincent et al. (ref. 6-18) used S192 and S191 data to study the feasibility of using ratios of spectral-band signals to map iron compounds in the exposed surfaces of rocks and soil and to differentiate silicate rock types. A general investigation of laboratory spectra had been.....

....conducted (ref. 6-19) by which the best bands for discriminating rock, mineral, and soil classes were determined. Tables 6-1 and 6-11 show the rankings of the simulated S192 bands for single-band and two-band ratio processing, respectively. The value of operating satellite sensors in narrow bands such as those employed in EREP is shown by the orderings. Using laboratory and field spectra, optimum spectral ratios were determined for mapping iron compounds. Comparison of the results of S190B photointerpretation and single-level slicing methods was also performed. Not all optimal ratios could be formed with the data set available, but those bands in the red, near infrared, and thermal infrared proved most useful. After suitable conversion and noise-reduction steps, three ratios were formed from the data.

The ratios R_8/7 and R_11/7 were chosen to differentiate ferric and ferrous materials. For proper implementation of the technique, the signal from atmospheric path radiance must be subtracted from the scene elements by using the darkest object subtraction method. Specifically, an analysis of the darkest objects in the scene is made, the mean value is calculated, and that value is subtracted from all pixels. After subtraction, the range of the ratio is adjusted for maximum contrast. For the test site shown in figure 6-27(a), the ratio image R_8/7 was constructed and is shown in figure 6-27(b). This ratio proved best for separating ferric, ferrous, and nonferrous classes of materials. The color codes define the range (low to high) of ratio values. The basaltic and acidic rocks are imaged in the test site as blue (low ratio), which is indicative of ferrous compounds. The....

.....high-ratio values (reds and oranges) relate to the ferric sediments that flank the slopes of the highlands. As the concentration of ferric iron decreases in the surficial deposits of the valley floors, the color shifts to yellow and green, an indication of intermediate values (fig. 6-27(b)).

The ratio process produced images that were much less influenced by variations in illumination and terrain shadows, a problem that reduced the effectiveness of single-band enhancement techniques. In one case, fault location related to the abruptness of a color-boundary change was enhanced.

The bands found useful in this investigation should help to define the optimum placement of bands for future multispectral scanner sensors. Current Landsat bands are excessively broad; Landsat band 7 integrates both bands used in the R_8/7 ratio.

Yarger and McCauley (ref. 6-7) found band ratios that gave best results for a statistical correlation with inorganic suspended solids in three Kansas reservoirs. The CCT's for the S192 data were used to select the proper pixels within each of the reservoir images for the nine bands analyzed. The pixels were averaged, and the signal level was converted to radiance levels.

The most effective bands were ratioed, and the values were plotted as a function of suspended solids (fig. 6-28). The red/green ratio improved the correlation with suspended-load when compared to individual-band performance. The ratio of infrared/green also exhibited a good linear correlation with suspended load. Agreement with Landsat data is also shown.

The effectiveness of the ratio method to discriminate between ice crystals and water droplet clouds was studied by Curran et al. (ref. 6-20) and Pitts et al. (ref. 6-21). These authors found that the infrared band centered at 1.65 µm ratioed with a lower wavelength band has potential for separating ice clouds, liquid water droplet clouds, and snowfields.

Techniques Involving Multiple Spectral Bands

Three or more spectral bands can be combined to produce a false-color image. In this way, the spectral bands invisible to the human eye can be rendered visible and used for interpretation of the scene. The generation of false-color images and their analysis was a major use of the Skylab multispectral scanner data. A second method for analysis of multiple-spectral-band data involved the use of pattern recognition techniques and digital computers. Many methods for computer-aided analysis of multispectral image data have been developed during the past decade, and these methods were applied to the Skylab multispectral scanner imagery [289] with generally successful results. Examples of the application of false-color imagery and computer-aided analysis of multispectral image data are summarized in paragraphs which follow.

In processing S192 data, the approach taken by Colwell et al. (ref. 6-11) was to determine the best combination of four channels that could be used to effectively classify an area. This approach was selected to minimize processing-time costs without sacrificing classification accuracy.

Each of the 22 scientific data output (SDO) channels (app. A, table A-l) of S192 scanner tape was inspected on a television monitor to screen out unusable data on the basis of noise level or saturation. Three channels were used in a color monitor to facilitate the selection of training fields. Coordinates of selected fields were determined from a grid network displayed simultaneously with the tape data. After selection of training sets, conventional statistical analysis of the data was performed.

The data were reclassified using the nearest neighbor algorithm as well as a threshold algorithm by which a point was reclassified only if a stated probability of correct classification was attained. In a preliminary test, SDO channels 2, 3, 8, and 12 were determined as necessary for classification; for another test, channels 2, 3, 8, 9, 10, and 12 were found to produce usable results. Later, for an extended area, nine channels (2, 6, 8, 9, 10, 12,18, 20, and 21) were used to determine the best combination of four channels. Classification tests made using channels 8, 18, 20, and 21 resulted in high accuracy, especially when the nearest neighbor algorithm was used. This result was expected because the nearest neighbor algorithm reclassifies each point as one of a defined set of crop classes, whereas the second classification technique using the threshold algorithm reclassifies a point as one of the defined set or defines it as unclassified. Further studies with EREP data yielded results based on analysis of all 22 channels for an area in northern Fresno, California, containing a variety of crop types in different states of maturation.

The best combination of four spectral bands for discriminating crop subclasses was determined to be 4, 7, 9, and 11, with band 10 or band 8 serving as a suitable substitute. However, the complexity of the agricultural scene investigated resulted in an overall performance accuracy of 57 percent. After reclassification into 13 classes for 22 subclasses, the performance accuracy increased to greater than 66 percent.

The studies showed that, depending on the agricultural environment, season, and state of maturity, the optimum combination of bands will vary. For a limited inventory of one or two crops, single bands might be adequate. The importance of the near-infrared region was particularly demonstrated.

Silva (ref. 6-12) analyzed S192 data for an Indiana test site by clustering of signal levels which showed that separation of 13 subclasses was possible. A separability measure was used to find the best 4 of the 12 bands evaluated; namely, bands 3, 7, 8, and 11. The four Skylab bands gave better results than four bands used to simulate Landsat bands, but this result could be due to both information and noise differences.

Overall, the performance of the S192 optimum four bands was better than the results obtained with digitized photographs and was essentially equal to results obtained from analysis of Landsat data taken the day before the Skylab pass, despite the fact that the Skylab data were noisier. Figure 2-11 in section 2 (Land Use and Cartography) shows the color-coded results for nine land use classes.

The classification procedure used the pattern recognition algorithms that have been implemented in a software package called LARSYS (ref. 6-22). Training fields for 12 classes, representing approximately 0.3 percent of the study area, were scattered throughout the study area.

The training fields were evaluated using a clustering routine to determine whether further division of the 12 classes was necessary to ensure that each of the resulting spectral classes represented a unimodal distribution. Assuming that the spectral data of the samples for each spectral class had normal (Gaussian) distribution, the N-dimensional mean vector and the N by N covariance matrix of the multispectral data sets were computed for each class, where N represents the number of channels of spectral information in each data set. The mean vectors and covariance matrices were used in the actual classification routine, which incorporates the maximum-likelihood decision rule (ref. 6-23) with the a priori probability of occurrence for each class being equal.

A modified divergence rule was used to select the best 4 of the 12 bands in the Skylab S192 data set evalu-....

....-ated, Divergence is used because it is a measure of separability between two density functions that represent two classes of objects. This modified divergence (called transformed divergence) was extended to a multiclass case to choose the best four bands in two separate ways.

First, the average of the transformed divergence for all possible class pairs was maximized; then, the minimum transformed divergence of all possible class pairs was maximized. It cannot be shown, however, that either of these methods is optimal. Four bands were selected for....

....the classification to reduce computer costs and to determine whether the bands selected included any spectral bands not available in the other data sets.

The products obtained in the study include the transformed divergence measures for the separability of each pair of classes in each data set, the classification maps, and the classification performance results. The classification performance results were obtained by selecting six test areas scattered throughout the study area representing the classes under consideration.

Photographs were projected onto the classification map of these test areas, and a point-by-point check of the classification was done. The points from all six test areas were combined to obtain performance results for each class and for the overall classification performance represented by the total number of points classified correctly divided by the total number of test-area points.

In the processing of S192 data by Goetz et al. (ref. 6-24), scan lines were added or deleted where necessary to correct for scale. Unusual lines were modified by inserting an average of neighboring lines and thereby reducing the noise level. The high-data-rate channels (1 to 16) were merged into a single band equivalent to the low-data-rate channels (17 to 22). Geometrical rectification was performed as a preliminary step to compensate for a 5.5° cone angle and for the 110° scan arc motion of the sensors. A correction was also made for spacecraft motion and Earth rotation during the acquisition time of a given frame. In the process, 1240 original samples along a scan line were reduced to 1056 samples/line.

Once in this form, images can be easily made to fill the dynamic range of the display medium by means of contrast enhancement. Pairs of images having intensities I₁ and I₂ were compared by a ratio algorithm of the form

(6-2)

where l_o is the output image intensity and the two constants a and b are chosen to maximize contrast and may change for different scenes. This technique proved useful for temporal and spectral comparisons.

One classification algorithm that was used is called stepwise linear-discriminant analysis (ref. 6-24). it consists of finding a transform that minimizes the ratio of the difference between group-multivariate means to the group-multivariate variances.

A second technique used was a hybrid approach. First, for each spectral band, the means and the standard deviations for each category are computed. The paired categories are compared for each band and determined to be separable if they satisfy the relation

(6-3)

where C is a constant. The hybrid classification combines two existing classification schemes, the parallelepiped algorithm and the Bayesian maximum-likelihood function.

The parallelepiped algorithm approximates a hyperellipsoid that is defined by computing means, variances, and covariances based on the assumption of a Gaussian distribution for the signal variations for each category in the selected wavelength bands. Decision boundaries, related to the number of standard deviations about the mean, can be defined by the computer operator. Each band is considered a vector component; [293] i.e., the set of spectral bands for a given category is a vector in multidimensional space. Ideally, all pixel vectors for a given category will cluster about a well-defined mean with small variances that fall within a narrow ellipsoid.

Maximum-likelihood processing depends on the knowledge of a priori probabilities. The data set can be integrated to define the probabilities, but the likelihood usually is defined by the multivariate Gaussian probability density function.

Goetz et al. (ref. 6-24) used an interactive unsupervised clustering algorithm. The interactive technique incorporates a self-grouping method based on the best partitioning of N objects into g groups by maximizing the variations between groups and minimizing the variations within groups. Several criteria then can be considered in selecting the best grouping. The methods using the various criteria are time consuming and require large-computer capabilities. Some random sampling of the scene to give the initial clustering is performed to make the technique more practical.

Other investigators, such as Sattinger et al. (ref. 6-25), in studying techniques for land-cover inventories, deleted the two doubly sampled thermal channels. First, noise levels and dynamic ranges were assessed by making a histogram of the data values in each spectral band. Some differences were noted in the histograms for even- and odd-numbered SDO's. By this process, only the better data channels were used.

An optimum band selection program was invoked, and the best six bands were determined. In order of preference, these were 0.78 to 0.88, 1.55 to 1.75, 0.98 to 1.08,0.68 to 0.76,0.52 to 0.56, and 0.62 to 0.67 µm. It appeared that the choice was dependent on the signal-to-noise ratio and on spectral contrast.

Seasonal comparisons using March to June data were considered best. By means of supervised classification techniques, the best single-band delineation (1.55 to 1.75 µm) was used as a base onto which 2.6-km² grid sections were transferred. The section lines had been traced from a 1:120 000-scale color-infrared transparency spatially registered to the digital map of S192 data by a transfer scope. This grid enabled better location reference for the training sets when analyzing the S192 digital data. Thirty-five separate sets were designated to encompass the wide variability of the categories of interest. Enlargements (8x) of S190B photographs proved valuable in selecting boundaries of training sets and judging the homogeneity of fields.

Means and variances for each training set were computed with all boundary pixels excluded from calculations. A test for signature statistical uniqueness was made by computing the probability of misclassification for all possible pairs of signatures. Each pair-wise probability of misclassification provides a measure of the separability between two multidimensional statistical distributions. It represents an average of the probabilities that samples from distribution A will be mistaken as B and that samples from distribution B will be mistaken as A. The results vary between zero (the two distributions well separated) and 0.5 (the two distributions superimposed). The classification rule used is the best linear decision rule used to classify multispectral data (ref. 6-25).

Based on the results of the pairwise calculation, the 35 signatures were aggregated into a smaller number of composite signatures by combining groups of signatures having high probabilities of misclassification. Classes defined by photointerpretation of high-altitude-aircraft photographs and composite signature analyses of S192 data can give different results. For example, for the scene classes identified on high-altitude-aircraft photographs taken in June 1972, there were three non-forested wetland classes, a flooded forest, brush fields, three types of forest crown cover, aspen regeneration sites, and a pine plantation. On the other hand, for the composite signature classes of the S192 data (acquired in August 1973) of the same scene, there were only two non-forested wetland classes and only one general forest class with the other classes remaining the same. Time and seasonal differences between the two sets of data could explain the different groupings. After the categories were selected, the optimum spectral bands were determined (table 6-III). Further studies comparing Landsat discrimination with selected S192 bands showed the significance of band 11 for improving performance as well as the value of using more channels (i.e., more than four).

Field checks were made to verify choices of training sets and their particular character. Finally, after redefining the training sets on the basis of onsite inspection, classification probabilities for each category were calculated. The expected classification performance for the best linear rule classifier was calculated.

Figure 6-29 illustrates the spectral separability of the 12 classes using the best 2 spectral bands. In this figure, the relative location, the shape, and the orientation of the distributions provide a graphic illustration of their....

....statistical uniqueness. The computer separated the classes using the six best bands, chose weighted probabilities for the scene (based on color-infrared photographic results), and determined a threshold value that corresponded to the 0.001 Ievel of rejection for six degrees of freedom. Pixels having values exceeding the threshold were unclassified. The accuracies of classification range from 72 to 82 percent, depending on the grouping of the final categories and the averaging of statistics over 2.6 km². Some limitations of accuracies are expected from the nature of the available signal-to-noise ratio, the misregistration of bands, the geometric distortion, and the scan-line-straightening procedures.

A technique for obtaining accurate crop area estimates in agricultural areas characterized by ground resolution sufficiently large to create a high probability that the scanner will integrate a mixture of objects was developed by Nalepka et al. (ref. 6-26). Figure 6-30 displays the nature of the problem over 2.5 km². In this example, the number of pure field pixels is only 30 percent of the total pixels that cover the scene. Thus, the impact of this technique can be significant, because the more mixture pixels in a scene, the greater the chance for error. Conventional classification techniques are not adequate for such a case.

For this technique, a small number of signatures having sufficient separation are desired not only to avoid degenerate signatures (one signature equaling a linear....

....combination of two others) but to keep processing time practical. Processing time is proportional to m(m + 1)/2, where m is the number of signatures.

In performing classification, signatures based on pixels are derived. For best results, center pixels or field pixels are more desirable for avoiding the ambiguities of border points and mixtures. Nalepka et al. (ref. 6-26) identified field pixels for training purposes by inscribing a small polygon within the boundaries of a training field.

In general, an inset I was defined that is calculated by

(6-4)

[295] where

indicates the scan direction x or the alongtrack direction y

is the size of the resolution cell in the direction of (The S192 data were oversampled by 10 percent; therefore, a digital resolution cell is not equal to a pixel.)

is the size of the pixel in the direction of .

B is the inset necessary to ensure that the pixel does not include the boundary between fields; typically, B = 0.5 pixel

is the error due to misregistration effects; e.g., if one channel is misregistered from the others by R pixels, this channel could still be imaging across the field boundary when the other channels are imaged entirely within the field. For conic data corrected for misregistration, R_x = 0.32 and R_y = 0. For scan-line-straightened data, R_x= 1 + M sin and R_y = 1 + M cos , wherein M is the maximum misregistration in conically scanned data (found to be 1.13) and is the angle between the line tangent to the conically scanned data at the point being considered and a line in the alongtrack or flight direction. To develop one measure for the entire scan line, the maximum values of sin and cos , which is 1, are used. Thus, R_x = 2.13 and R_y = 2.13

L is the error due to field location errors

S is the error due to movement of individual pixels as a result of the nearest neighbor scan-line straightening. Therefore, for conically scanned data, S = 0; for straightened data, S = 0.5 pixel

Thus, the inset to be used for conically scanned data would be

(6-5)

The inset to be used for scan-line-straightened data would be

(6-6)

where l = l_x = I_y

This technique was applied to an urban scene to determine the amount of vegetative and impervious material. Such information is useful to geographers, urban planners, and urban climatologists. Most pixels could be mixtures. Five classes of interest were defined: green vegetation, concrete, other impervious materials....

[296] ...(roofs, asphalt, etc.), bare soil, and water. A statistical test for a measure of separateness (approximately the distance, in standard deviations, between the signature mean and the hyperplane through the other signature means) showed that the five classes were degenerate. Data with limited signal ranges and less than optimum spectral contrasts are not suitable for mixture processors.

Results of this investigation indicate that, because of misregistration in line-straightened data, processing should be performed on a conical format. Finer spatial resolution should be considered for future sensors to reduce the number of field pixels for better training signatures. Also, the design of future sensors should incorporate a means of adjusting scanner gain and offset parameters to match the radiance characteristic of individual scenes better so that the available dynamic range is used. Long atmospheric paths add a sizable constant radiance and attenuate the reflectance radiance to cause reduced contrast.

Signature extension is a process by which the training statistics from one scene may be modified and used to classify features in a second scene that may differ geographically or temporally. Use of this process reduces the need for extensive ground truth and for retraining signatures without incurring an intolerable loss in accuracy. Nalepka tested several signature-extension algorithms on the S192 data (ref. 6-26). The data were grouped into 10 clusters that were identified from training statistics. One of the first methods attempted involved dark-object signature correction. The technique assumes, band by band, that the signal levels generated by dark objects represent mostly path radiance and can be used to generate a correction algorithm. Because of the haze, the correction factor tends to be greater as the wavelength becomes shorter. Significant improvement in classification accuracy occurred; in particular, the recognition was unexpectedly accurate where the haze was densest (i.e., the dark level correction was greatest). However, the algorithm showed some tendency to misclassify bright features as darker features. The need for a multiplicative signature correction in conjunction with an additive correction feature was made evident.

A second method, involving an adjustment for the mean level of the signal over an extended scene, was tried and produced slightly less successful results. In the technique, the correlation between averages over portions of the training scene and of the signature-extension scene is used to estimate a correction for the mean levels of each training signature in each band. The method required similar percentages of ground cover. The results showed nearly the same correction factor for the shorter wavelength bands, although differences were noted at the longer wavelength bands. This result suggests a bias in favor of darker materials, which absorb shorter wavelengths; at longer wavelengths, the scene has more contrast and vegetation is more reflective.

A third method incorporated both a multiplicative and additive signature correction (MASC) factor by using a least squares regression to match training cluster mean signal levels with local cluster mean levels based on the ordering and spacing of those signature means within a chosen data band. Because mathematical modeling of the illumination variations indicated that the variations should be both multiplicative and additive, use of MASC was expected to produce a realistic signature correction factor. A test performed using only the 1.55- to 1.75-µm band for classifying a Michigan agricultural test site showed that more than one band should be used to help match the clusters from differing portions of an extended scene. When there are variations alongtrack, clustering algorithms cannot be expected to produce sets of signatures from two different scenes that are in close correspondence; some method is needed to identify and omit noncorrelating clusters during the cluster-matching procedure that leads to calculations of signature correction factors.

Simonett (ref. 6-27) used the S192 data to extract land use information. Contrast-enhanced, geometrically rectified gray-level maps for the test areas were printed for the spectral bands (4 and 8) providing the greatest visual discrimination of land use (cover) categories. Noise-reduced, line-straightened data were used and reformatted for further analysis. Contrast enhancement was accomplished by a technique called histogram equalization (ref. 6-28), in which a nonlinear (several to [297] one) mapping of input gray levels to output gray levels is defined and the input data are effectively spread across the entire usable dynamic range of the output display. Because only land use was of interest, only the range pertinent to land features in a given spectral band was transformed. The output print matched the scale of U.S. Geological Survey (USGS) 7.5' quadrangle maps for ease in comparing results with ground truth and in selecting computer training sites.

Finally, ground-truth and scanner digital data were merged by generating a data file consisting of map information on field boundaries, land use boundaries, and 13 spectral data values. Thus, within each individual record of the data file, the analyst had access to the gray levels from all S192 bands, the pixels that were assigned to known fields, and the land use represented by the pixels. The 22 SDO's were reduced to the best 13 spectral bands based on a selection process.

Field signatures were computed to give the means and the covariance matrix, and a signature file was generated. By means of a stepwise discriminant analysis technique (refs. 6-29 and 6-30), the best set of spectral bands for discriminating various land use categories was found. Composite group signatures (a function of how a given scene is to be divided) were used to calculate a set of cross-product terms to form a matrix for both within group cross products and total cross products. The total cross-product matrix was directly proportional to the variance-covariance matrix for all the data treated as a single data set. The within-group cross-product matrix was directly proportional to a weighted sum of the variance-covariance matrix for each group. A given spectral band provides a good discriminant between groups if the total variance for all the data (diagonal element of the total cross-product matrix) is much greater than the variance obtained by treating the data in groups (diagonal element of the within-group cross-product matrix). Spectral values are assumed to have a multivariate Gaussian distribution throughout. The best spectral band for discriminating among all groups was selected by calculating the likelihood ratio to test the equality over all groups for each spectral band. The process is repeated to find successive combinations of bands that discriminate best among groups, given that specific bands have already been selected. In applying the stepwise discriminant analysis, only the best field signature representing the group selected for analysis was used. In addition, an estimate of the spectral separability of the input data classes at each step was made using the transformed divergence measure (ref. 6-31). The technique was used to examine 87 test sites with signatures calculated for 609 fields and to separate them into relatively broad land use classes (Level 1)-urban, agricultural, forest, water, and wetlands-and more specific land uses (Level II).

For the broad classification, five bands were found to be important. These bands were 11, 9,13, 5, and 6, in order of decreasing value. Again, spectral coverage in the near infrared, together with the red and thermal-infrared bands, proved to be crucial.

Further examination to determine separability within each broad class was undertaken. However, the similarity of the spectral signatures for Level I classes, the complexity of the spatial distribution, the lack of multidate coverage, the presence of noise, misregistration, and different sampling rates between channels influenced the final outcome. The best results were obtained for large fields, in which some of the effects were not as important. In general, the best five bands for discriminating the general classes were not the same bands for discriminating elements within a class; therefore, a two-stage processing technique was implemented.

In the first stage, an unknown pixel signature was assigned to one of five general levels and that information was stored. In the second stage, the best five spectral bands for discriminating within one of the five general levels were selected. A maximum-likelihood classification algorithm was used during each of the two stages with equal a priori probabilities assumed for each class.

Natural grouping of signal levels, or clustering, was also tested. The distance measure wed in the clustering is the reduction of the divergence measure for an assumed diagonal covariance matrix. Use of this procedure reduced computer-processing time but resulted in some loss of discrimination. The investigation revealed the value of multidate imagery because of the character of outlying elements within certain groups.

[298] Thresholding within two standard deviations of the mean was performed only during the second stage. To check classification accuracy, all fields having signatures were stratified as either training or test fields. The results varied from an overall accuracy greater than 70 percent for Level I separation, with forests classified at 84 percent and water at 96 percent. Thresholding reduced the number of misclassified pixels.

Two important conclusions were reached concerning the improvement of classification accuracies. These were the need for multidate coverage to help in the spectral separability of classes and the need for higher spatial resolution in proper registration to reduce the problems of mixed pixels.

Hoffer (ref. 6-32) performed a computer analysis on S192 data taken over mountainous terrain and classified the scenes for a variety of applications. As an initial step before computer classification, the data were geometrically corrected and made compatible with Landsat, aircraft, and digital topographic data. The digital topographic data were used to obtain elevation information, and an interpolation technique was developed to compute slope and aspect (fig. 6-31). All data sets were matched to an X-Y grid base and stored on a single computer tape for ease in multisensor comparisons. Twenty channels were available for further analysis; i.e., 13 bands of the S192, 4 bands of the Landsat multispectral scanner, and 3 channels of ground information (elevation, aspect, and slope). The information was geometrically corrected to a 1:24 000 scale to match the USGS 7.5' quadrangle topographic maps.

Two different processing techniques were developed and applied. The modified clustering technique provided generally better results than did either the standard supervised procedures or the standard clustering procedures used previously. Machine time was significantly reduced, classification accuracy was increased, and man/machine interactions were improved.

The modified clustering technique is essentially a combination of the supervised and nonsupervised classification procedures. It is designed to overcome disadvantages inherent in both approaches. The modified clustering technique is performed in four steps.

---

1. Several training areas of approximately 1600 pixels each are selected over the entire study site to be classified.

 

2. Each training area is clustered separately. The results are compared to a supporting data map for accuracy, and reclustering is performed if necessary.

 

3. The results from all training areas are combined using the separability algorithm, and a single set of training statistics for the chosen categories is developed.

 

4. The training area is classified again as a preliminary check, statistical parameters are modified if necessary, and only then is the entire study site classified.

 

A newly developed algorithm, extraction and classification of homogeneous objects (ECHO), provided more accurate classification in a more useful format than did conventional per-point classification. This classifier uses an algorithm that first defines the boundary around an area of similar spectral characteristics....

....and then classifies the area within the boundary as a single spectral class. The classifier specifies the boundaries as part of the procedure and is not dependent on the analyst for specifying boundaries. The output format proved acceptable to those agency personnel interested in a generalized cover-type map. The classification results of the two approaches are shown in figure 6-32. The per-point classifier results do not have the smoothing effect shown by the ECHO classifier results.

Overall evaluation of computer processing revealed that four-band classification was optimal from the standpoint of best accuracy and least expense; a significant increase in classification costs resulted from an increase in the number of bands beyond four. However, for different classes of objects, various combinations of four wavelength bands were needed for accurate discrimination. Thus, future satellite sensors should operate with more than four bands but need to analyze with only a subset of four bands, depending on the scene composition. For example, in analyzing a land use (forest cover) scene, 85 percent overall classification accuracy was achieved using the four bands located at 0.46 to 0.51, 0.78 to 0.88,1.09 to 1.19, and 1.55 to 1.75µm.

Acreage estimates of forest cover types were highly correlated with measurements from standard photointerpretation techniques. The value of the technique in [300] which slope and aspect were monitored together with the spectral signature was proved when spectral differences between forest cover types were found to be influenced significantly by topographic relationships, whereas spectral variations within individual forest types were related to differences in standard density. The consideration of these factors resulted in a substantial improvement (10 percentage points) in classification performance.

An evaluation of the priority for the various bands supported the use of six bands (spaced two in the visible, two in the near infrared, one in the middle infrared, and one in the thermal infrared) for the application area defined as land use and forest cover. When the techniques were applied to the mapping of hydrological features, the different classes of snow were tabulated as a function of elevation and spectral signature. This class difference arose from various mixtures of snow and forest cover within each scanner resolution element and possibly from snow-melting conditions.

The "layered classifier" approach was used and provided better differentiation between snow and clouds and better snow-cover class tabulation by comparison to conventional maximum-likelihood classification techniques. The thermal-infrared band proved valuable in measuring reservoir temperatures accurately after a two-point nonlinear calibration technique was used to process the S192 data. The high elevation of the reservoir and the transmission characteristics of the 10.2- to 12.5-µm band led to close agreement between satellite measurements and reference measurements for the same portion of the reservoir.

A tabulation of the watershed area extent of each spectral class of snow indicates that remote-sensing data can be effectively used to predict water runoff from mountain snowpack areas. On a regional basis, this information would aid reservoir and watershed management planners.

Thomson (ref. 6-33) investigated the effects of variation in the atmospheric state on pattern recognition performance. The question of whether preprocessing must be altered along a flight track is crucial to the success of signature extension. Based on a comparison of results obtained using a combination of four-band optimum sets, a Landsat analog, and optimum seven-band sets from the S192 sensor, the following conclusions are noted.

At atmospheric conditions equivalent to horizontal visibilities of approximately 10 km, adequate performance (greater than 65 percent correct classification) can be achieved within visibility variations of ±3 to ±4 km. If greater variations are found, data-preprocessing corrections must be made to maintain adequate performance. As better sensors are constructed, more care must be taken in selecting a set of bands and in determining their number and location. For example, although seven optimum bands gave better performance under noise-free conditions for a given set of environmental factors than did four optimum bands, the seven-band results were more sensitive to changes in atmospheric visibility.

In analysis of S192 data, Wiegand et al. (ref. 6-34) used the 0.78- to 0.88-µm band to visually differentiate water, vegetation, and bare soil on a cathode-ray tube in a study of eight saline areas in southern Texas. Simple linear correlation analyses were used to relate field electrical conductivity EC_e measurements to the mean multispectral scanner digital values for both bare soils and vegetation test areas. In one test area, bands 6 to 11 of the multispectral scanner correlated well, but the difference between the signal from bare soil and vegetation in the 1.2- to 1.3-µm band (in Landsat, 0.8 to 1.1 µm) correlated best with the amount of electrical conductivity or salinity.

Hannah et al. (ref. 6-35) used S192 CCT's to obtain land use information for Orlando and Lakeland, Florida. Band 13 imagery indicated that commercial-industrial regions, newly formed residential areas, and wooded residential areas could be separated on the basis of temperature. The land use maps prepared for these urban areas were derived from analysis of bands 4,6,11 or 12, and 13. The maximum-likelihood procedure was used principally, although less accurate schemes for classification were applied. These schemes were based on measurements of least distances from the point being classified and on determination of the three....

.....nearest classes by least distance and use of maximum likelihood to choose among them.

In a different approach, McMurtry and Petersen (ref. 6-36) used wave-number analysis. Each band inspected in the Fourier transform domain was found to provide unique wave-number information about particular scenes. This method is suggested for studying lineaments and geologic structures such as folds and fracture traces.

If the lineaments are related to fracture traces and appear as a line of discontinuous spikes in intensity relative to the surrounding intensity, they will have the distinctive wave number k (k = 1/, where is spatial wavelength) response for a lineament from k = 0 to the Nyquist wave number. Several filter functions can also be used with this method. For geology, a filter function called the strike-selective filter can be used to check for lineaments first observed by photointerpretation. If the lineaments exist, the digital format can be enhanced as shown in reference 6-36. Other features can also be enhanced by using this method (ref. 6-37).

Functions that take the derivative of the signal such as high- and low-pass frequency filters can also be applied. Filter analysis can aid in deciding which bands to use for the analysis of a particular parameter and can aid in selecting a subset of bands for color displays.

In a technique first outlined by Horwitz et al. (ref. 6-38), a proportion estimation algorithm is used to estimate the percentage of area occupied by different objects within the field of view (FOV). Geometrically, three objects as seen in two spectral bands can be depicted (fig. 6-33). Points A, B, and C represent the pure spectra of each object. If an instantaneous FOV contains a mixture of all three materials, then the signature X must lie within the triangle formed by connecting the vertices at A, B, and C. By extending a line from one vertex through the unknown point within the triangle and intersecting the opposite side, an estimate of the pairwise proportion of the pure materials constituting the unknown element can be made by taking the inverse ratio of the lengths into which the leg is divided.

The concept can be extended into N-dimensional space so that at least N-1 spectral bands of information are required to estimate mixtures of N objects satisfactorily. Some degradation in estimates is expected if one material is very similar to the weighted average of the others.

The value of the S192 spectral resolution was demonstrated by Poulton and Welch (ref. 6-39), in the study of rice-yield prediction. The crop calendar for rice begins with a bare unflooded stage, progresses to a fully green vegetative state, and eventually becomes a changing yellow during ripening. Thus, spectral discrimination at the proper season was found to be a dominant factor. For a test area in Louisiana, interpretation of an S192 color composite yielded a higher accuracy than did analysis of the S19OA and S19OB photographs taken at the same time. Figure 6-34 is a comparison of the three images. The Skylab S192 color composite using bands 1, 7, and 9 had better contrast and thus aided in the differentiation of rice in various stages of maturity from other crops when the field size was greater than 6 to 8 hm².

In geology, Goetz et al. (ref. 6-24) separated different lithologic units most effectively by using bands outside the range available in the current Landsat regions. They found that bands 0.46 to 0.51, 0.98 to 1.08, and 1.20 to 1.30 µm from both field spectral analysis and enhanced color composites gave the best separation (fig. 6-35). For this particular site, thematic maps produced by three classification algorithms were not as accurate as results obtained by photointerpretation of computer-enhanced imagery. Among the three algorithms tried, the linear-discriminant algorithm seemed to be the most efficient. The need for operational satellites capable of imaging in the far-reflective-infrared region was expressed.

Houston et al. (ref. 6 40) investigated the potential of S192 data by using both visual qualitative examination and densitometric quantitative measurements. Except for the detection of red beds in band 2 (0.46 to 0.51 µm) and better contrast of small, closed anticlines in band 4 (0.56 to 0.61 µm), the near-infrared bands provided the best contrast for different rock units (band 8, 0.98 to 1.08 µm; and band 9, 1.09 to 1.19µm). Relative density values for 15 lithologic units on each S192 band were measured by a video densitometer, and results showed that bands 7, 8, and 12 yielded the highest total contrast values. Computer analysis of the S192 data for geologic applications confirmed the value of the near infrared in enhancing contrast between lithologic units. The digital....

[306] ....tapes for the 13 S192 bands were analyzed by constructing alphanumeric maps and computing pairwise clusters from two-dimensional frequency histograms of test sites. Histograms for each of 78 possible pairs from 13 bands were analyzed, and maps were made of the best contrast pairs. A further analysis of reflectance vectors for the test region resulted in a decision to use weighting factors. One particular method, designated Q-mode factor analysis, was conducted with more than 100 000 points, and results indicate that this method is useful. If matrix D is an m by m (m = n) data matrix of reflectance values in the i-th channel for the j-th pixel, the columns of matrix D will be normalized to unit length L by the formula

(6-7)

representing the square root of the sums of each matrix element squared, and by the new matrix

(6-8)

Each column of Z_Q represents the reflectance at a given location after its brightness has been adjusted to the same general intensity.

By taking the cosine of the angle between two normalized reflectance vectors, a test of similarity is easily made. The cosine will have a value of 1 if the vectors are identical and will have a value of 0 if the vectors are different (perpendicularly). A computer algorithm was developed that facilitated factor comparison (ref. 6-40) of large numbers of pixels by Q-mode analysis. Table 6-lV is an example of the factors computed by the Q-mode method for varying wavelength . The use of the factors in making a new alphanumeric map shows the manner in which outcrops of enriched rocks could be enhanced or the densest vegetated areas could be delineated.

Haefner (ref. 6-41) found that, for operational mapping of dynamic features such as snow, preference should be given to digital data to take advantage of near-real-time classification. By combining information from band 11 (1.55 to 1.75 µm), band 7 (0.78 to 0.88 µm), and band 2 (0.46 to 0.51 µm), snow could be separated from clouds. The test area was first classified into 29 different categories but later reduced to 5 main classes: snow in Sun, snow in shadow, clouds, snow-free area in Sun, and snow-free area in shadow. Data were prepared from digital tapes following the standard procedures of (1) data reformatting to match available computer systems, (2) delineation, (3) statistical evaluation of sampling areas, (4) classification of classes based on euclidean distance, and (5) data display with geometric corrections for maplike output.

Goldman and Horvath (ref. 6-42) researched the detection of oilspills on the ocean with S192 data, which were expected to be useful for this application because of the spectral resolution choices and the wide area of effective coverage (68.5 km). Difficulties were expected because of the uncertainty in the nature of oil-to-background contrast. Slicks that have a center thick enough to inhibit the diffuse upwelling radiation from the water would appear darker than the surrounding water. Because oil has a higher specular reflectance than that of water, the water may be dark and the specular reflectance of the oil may dominate; in such cases, the oil would appear bright in contrast against the background. Chlorophyll and suspended particles complicate the spectral contrast further. In almost all cases, however, for moderate-thickness films, the reflectivity of at least the thin portions of oil on water should be uniformly higher than that of water alone. Techniques for ratioing total radiance values from one band to another can help separate these effects and perhaps be used to confirm the location of an oilspill.

Statistical analysis of the noise in the S192 data was performed before the detection method was attempted. Mean values and standard deviations were computed for a test area. The results showed that a small reflective anomaly could not be detected with single-band analysis only.

Spatially coherent weighted sums were generated to reduce noise effects using four channels (SDO's 3, 7, 9, and 15). Weighted summations, incorporating ratios of the standard deviation, equalized the noise contribution.....

....of all channels used in the summation. Values of the sums falling within the highest and lowest 10 percent of the range were isolated and designated special points, if an oilspill were present in the scene analyzed, it would, with high probability, have special points adjacent to each other.

Because the results over a 600-pixel area gave only a random distribution, no oilspill could be confirmed. However, the technique proved valuable because confirmation of an oilspill with Landsat bands 4 and 6 was successful when ratios of radiance were taken after subtracting the background radiance from all values in the scene.

Pirie and Steller (ref. 6-2) studied coastal circulation and sediment loading using S192 color composites of images specially prepared by a linear expansion of the image density range so that contrast was enhanced for small density changes. Bands 4, 6, and 7 were used to enhance interpretation in analysis of coastal processes (fig. 6-36). The detail seen in the enhancement is considered unique. Variations in suspended-sediment load are easily seen, and comparisons to known sediment distribution maps show close correlation. Areas on the enhancement that appear to be receiving the greatest amounts of surface sediment closely correspond to the areas of maximum deposit as seen from ground-sampled surveys.

Water-depth relations were established from analysis of S192 visible band imagery by Trumbull (ref. 6-43) and by Polcyn and Lyzenga (ref. 6-44). Band 3 (0.52 to 0.56 µm) was most useful for water penetration; color coding by Polcyn improved contrast which aided in relating data values and water depth. Further discussions are presented in section 5.

Imagery Use in Earth Resource Models

The 13 S192 bands have provided investigators a wide range of spectral data for study of Earth resources with predictive models. The examples discussed in this section illustrate the use of S192 imagery in models that are used to predict climate changes, wind fields, atmospheric effects, and soil moisture.

Alexander et al. ² used the thermal band of the S192 to construct observed-temperature maps for comparison with predicted-temperature maps produced by use of a surface-climate simulation model. By considering the effect of land-use-related components of urban climates (such as the heat-island effect), better information on climatological consequences of land use changes can be derived from the model and future urban design can be made more effective. This model enables extrapolation of remote-sensing results at a given time to other times of the day or year while allowing for changes in the input parameters. Data obtained by an imaging radiometer enable construction of a matrix of spatial averages over all types of urban surfaces that is far superior to point-sampling tables. Earlier research had demonstrated the value of time-sequential remotely sensed data. The simulation model is based on the familiar energy conservation equation relating

where net radiation R, soil heat flux S, sensible heat flux H, and latent heat flux L are expressed in terms of meteorological-geographical parameters and surface temperature. A temperature equilibrium model is used to search for the specific surface temperature that balances the equation. Profiles of soil temperature as a function of depth are updated after each iteration. The Skylab EREP experiment led to improvements in a second version of the model that incorporated estimates of geographical terrain parameters in the form of (1) wetness fraction (irrigated lawn cover and tree cover), (2) silhouette ratio (the ratio of the vertical silhouette area in a tract to the horizontal area of that tract), and (3) observation height (mean vertical height that obstructs airflow and also increases the area of absorbed radiation). These factors were even more valuable because they could be deduced from remotely sensed imagery.

[309] From space, the temperatures defined for a pixel can be ambiguous when several objects are in the same resolution element. To calibrate the scanner, a method is used to account for the atmospheric path influence on surface-temperature estimates. A gray-window model (ref. 6-45) was used in the form

(6-10)

where R_z is the radiance in watts per centimeter squared micrometer steradian received by the spaceborne scanner, R_o is the surface radiance, E is the emissivity of the airpath, and [E_bb()] is the equivalent black-body radiance of the airpath at mean temperature .

Both atmospheric path convections and two-target temperature-calibration procedures were used to finally relate scanner digital values to true surface temperatures. The results were encouraging. However, when the model prediction and S192 temperature maps were compared, the model produced values lower than those observed. Further refinements of the model and better classification of land cover are recommended. The use of computer-derived classification from multispectral scanner data was considered a promising alternative.

Villevieille and Weiller's (ref. 6-46) use of models concerned the evaluation of vertical wind profiles. The convective cell in the atmosphere can be considered a stationary phenomenon that is spatially repetitive with an interior sinusoidal vertical wind variation along the vertical axis and two perpendicular horizontal axes.

One example of a wind field W would be

where I = 2 /L_x, m = 2/L_y, and n = 2/L_z define the wave numbers of wavelengths L_x, L_y and L_z, respectively, that can be measured from satellite photographs. In the presence of "cloud streets," L_x is the spacing between the centers of two consecutive clouds in a single street, L_y is twice the spacing between the axes of two streets, and L_z is twice the thickness of the convective layer.

Several different cell types were studied by means of a method in which numerical techniques were used, friction forces were considered, and wind profiles at the interior of the convective layer were obtained for a given characteristic dimension. The method was applied using S190B data obtained over Bordeaux, France, to determine spacings L_x and L_y. Values of other parameters needed were obtained from radiosonde measurements. Figure 6-37 shows a comparison of the predicted profile and the actual profile as measured by radiosonde.

The cell type determined the minimum interval energy dissipation. The results warrant further comparisons. If successful, interstreet spacing and street directions derived from satellite measurements could be coupled with mean thermal gradients derived from surface temperatures supplied by a very high resolution radiometer or other future satellite systems to make feasible the operational use of the model.

Aldrich et al. (ref. 6-47) used an aircraft to obtain satellite-matched forest terrain-reflectance measurements so that corrections for atmospheric effects could be made. The aircraft platform afforded more versatility than would a tower-mounted instrument, which can only measure signatures over an area equivalent to a few pixels of satellite data. With multidate coverage, the possibility of measuring time changes in vegetative spectral signatures was feasible. Also, information of importance for signature-extension techniques in computer-aided classification schemes could be obtained.

The aircraft instrumentation consisted of am upward-pointing irradiance meter and a downward-pointing radiometer. In this instrument, silicon diode detectors were filtered to match the S190 sensors and the Landsat-l multispectral scanner bands.

A video camera monitored the flight track so that targets actually measured could be verified in real time. At low altitude, which minimizes effects of the atmospheric path, the aircraft radiance N_a is

(6-12)

where is reflectance, H is irradiance, and Lambertian reflection is assumed. Satellite radiance N_s can be equated to the sum of two effects: the path radiance N_p, and a term proportional to surface reflectance by the....

.....product of total irradiance and atmospheric transmittance .

(6-13)

The FOV of the instrument is 2.6°; therefore, at a 300-m altitude, the ground resolution is 13 m. The output was recorded on an airborne chart recorder and later sampled at 20-m intervals. Calibration was performed in a laboratory using National Bureau of Standards traceable standards. Ratios of radiance to irradiance were calculated from these values for areas that could be related to satellite pixels. Skylab S19OA photographs were scanned with a digital microdensitometer. Density was calibrated by a comparison of measured duplicate densities produced by the same exposures applied to the original film. After suitable conversion from digital satellite counts to diffuse density and relative exposure, an equivalent radiance N was computed from the relation

(6-14)

where E is absolute exposure; F is the camera lens f-number; t is integrated exposure time; and T is total transmittance of lens, filters, and window.

Thus, by measuring Skylab data for N_s and using the low-altitude reflectance measurements for for at least three different ground reflectors, estimates of N_p and can be derived and used for correcting atmospheric effects. At satellite altitudes, total irradiance is assumed to be that of solar input.

[311] In a plot of N_s as a function of , the intercept on the radiance axis gives N_p, whereas the slope is proportional to beam transmittance and irradiance. When the correct values are known, new radiance values can be corrected and adjusted to earlier radiances derived from scene classes of the same reflectance.

In a Skylab EREP investigation (ref. 6-48), a method developed by Colwell (ref. 6-49) was applied to the S192 sensor data to determine surface soil moisture in the presence of partial vegetation cover. The reflectance of bare soil decreases for increasing soil moisture (fig. 6-38). The difference in reflectance between wet and dry soils is greater in the reflective infrared than in the visible part of the spectrum. Also, as the amount of green vegetation cover increases, the near-infrared (0.7 to 1.1 µm) reflectance increases but red reflectance decreases because of chlorophyll absorption. If soil moisture is to be inferred from a field with partial vegetation cover, some means for removing the effects of vegetation cover independent of the effects of variable soil moisture is needed. Fortunately, the near-infrared/red reflectance ratio is more sensitive to the amount of vegetation cover than is either of the single bands. In addition, the ratio is insensitive to soil moisture. Figure 6-39 summarizes the important relationship that suggests the method will work to normalize the effects of both soil type and soil moisture. From these trends, an algorithm was developed in the form

(6-15)

where A', B', C', D', and E' are constants; , is infrared reflectance; and is red reflectance. The term is a correction for the nonlinearity of the effects of both soil moisture and vegetation, and the term is a correction for the nonlinearity of the relationship between percent cover and the infrared/red reflectance ratio.

A specific relation was formed after simulating the reflectance of canopies with 30 combinations of soil moisture and vegetation cover and after finding the....

.....standard least squares regression relation. The specific equation became

(6-16)

The correlation index R² of 0.95 for the values used in this algorithm indicates the effectiveness of the algorithm in correcting for the effects of variable vegetation cover and in accurately predicting the surface soil moisture. Such modeling can be useful in determining sensor requirements (spectral bands and signal to-noise characteristics) for future satellite systems in specific applications. For example, the modeling suggests that a noise equivalent reflectance difference NE of 0.5 percent or more might be required in the near-infrared spectral band for estimating surface soil moisture in variably vegetated terrain.

[313] The advantage of a remote-sensing technique for the determination of surface soil moisture under variable vegetation-cover terrain is the ability to map large areas and to locate homogeneous areas with different surface soil moisture regimes. When this technique was applied to aircraft multispectral data similar to the S192 imagery, positive results were obtained. In figure 6-40, a portion of the test site that was mapped is shown; the map gives the areal distribution for soil moisture. Techniques using only one band failed to show correlation with soil moisture. Under certain conditions, the use of the ratio infrared/red reflectance, together with the scanner thermal data, will improve the significance of the regression equation for soil moisture.

Because of better registration between bands, S192 conical data were used to construct a land-cover map of an area in southeastern Ontario, Canada. A particular feature of this map was the 10 classification categories shown in figure 6-41. These classes (water; marsh; mixed conifer/hardwood; hardwood; suburban; quarry and bare soil; undifferentiated herbaceous; and low-, medium-, and high-percent green herbaceous cover) are suggested as being pertinent to hydrological problems dealing with runoff, water balance, and water management applications. The map was made in two key steps. First, an unsupervised clustering algorithm was used on six polygon-shaped test sites chosen from aircraft underflight imagery 1 day after the Skylab pass. Second, the reflectance ratio R_10/6 was used to determine the percent cover of green vegetation. The final classification was performed using 30 signatures to define 8 classes. Water was also recognized as a level slice from the near-infrared band. Probabilities of correct classification were computed using a program to generate 1000 points for each signature with a normal distribution. A best linear classification rule was used to obtain correct probabilities of classification that were greater than 90 percent correct for six of the eight classes and approximately 77 percent correct for two classes, which were suburban and undifferentiated vegetation. This result was expected because of the mixed nature of the class defined.

Comparison with Landsat recognition classes for the same area showed similar results for 75 percent of the scene. The maps were reasonably equivalent in their information content. Most of the differences could be accounted for by differences in training-set signatures used.

An extensive modeling program to use S192 imagery for delineation of optimum fishing areas in the Gulf of Mexico was performed by Savastano (ref. 6-50). Detailed discussion of the model is presented in section 5.

THE S191 INFRARED SPECTROMETER

Although the S191 infrared Spectrometer was not an imaging device, it was designed to obtain critical information about the spectral transmission of the atmosphere and about reflection characteristics of ground classes of terrain important to Earth resources applications. This knowledge is useful directly in the study of atmospheric processes and in the design of multispectral scanners such as the S192. The search for the optimum spectral bands for a given application is one of the major problems in remote-sensing research.

The S191, providing continuous spectra over a wide range, was manually pointed at selected targets to obtain the spectral characteristics of represented areas on the surface. Data in two ranges were obtained: between 0.4 µm in the violet and 2.5 µm in the near infrared, and between 6.6 and 16.0 µm in the thermal infrared. Scan time was I second, and the ground area coverage was approximately 500 m in diameter.

Data were collected in six spectral segments across the two ranges (appendix A); with appropriate reduction and calibration procedures, the data were reduced to a form suitable for analysis. Many Skylab investigators used the sensor for atmospheric radiative transfer experiments, whereas others found it useful in measuring the surface radiance more accurately.

Anding and Walker (ref. 6-51) integrated the radiance values from the two portions of the thermal band on each side of the ozone absorption band and ratioed the values to offset the effect of atmospheric influences on sea-surface temperature measurements. This technique of using two thermal bands such as those available from the S191 infrared Spectrometer offers an attractive alternative to correcting radiance values for atmospheric temperatures, moisture, and aerosol content. Further discussion of Anding and Walker's investigation is found in section 5.

Silva (ref. 6-12) used the output data in the thermal infrared to compare the data on spectral radiance measured by the S191 with data derived from surface measurements using both a ground spectroradiometer and a pyrheliometer to determine the spectral radiance and the spectral atmospheric transmission. The spectral path radiance, as computed from S191 measurements and as predicted by atmospheric models (ref. 6-52), is [315] compared in figure 6-42 for the visibility condition present at the test site (Lake Monroe, Indiana). The temperature of the lake was measured, and figure 6-43 shows the comparison of the spectral radiance derived from the S191 and that due to a black body at the temperature measured. Agreement of the results within experimental error verified both the atmospheric model and the feasibility of detailed spaceborne spectroradiance measurement.

To derive quantitative measurements of stratospheric aerosol characteristics, Tingey and Potter (ref. 6-53) used S191 data in the 0.4 to 2.5-µm band to acquire high-spectral-resolution data for increments of 2.4 km of altitude at the Earth limb. As an initial step, some signal averaging was introduced to improve the signal-to-noise ratio at the low radiance value encountered. Compared with the S192, the S191 proved to be more sensitive but less accurate in its absolute radiometric calibration. However, analysis of the S191 data proved that aerosol layers could be detected in several spectral bands at various altitudes measured from the top of the atmosphere. It was found that layers at altitudes of 42, 50, and 55 km are more responsive to longer wavelengths, whereas layers at 59 and 66 km were more responsive to wavelengths near 0.53 µm. Knowledge of the distribution of aerosols is of interest in astronomy, meteorology, air-quality surveys, and remote sensing.

In geology, Vincent et al. (ref. 6-18) applied S191 data to the problem of differentiating basaltic rocks from dacite. Atmospheric models from Anding et al. (ref......

---

---

.....6-54) and radiance data at Yucca Flat, Nevada, were used to calculate the average spectral radiance at the surface. To compute spectral emissivities for basalt and dacite, corrections for the different temperatures of the two materials were made. The ratios of the two emissivities were calculated by averaging emissivity in the short-wavelength band and dividing by the average emissivity in the long-wavelength band. This calculation was accomplished for different bandwidths pro. posed for future space sensors; in each case, however, the emissivity ratios for basalt and dacite were different, an indication that the materials could be separated by means of ratio techniques operating in the thermal-infrared bands.

In oceanography, Maul et al. (ref. 6-16) used the S191 nadir data to develop methods for recovering the ocean color spectrum through the atmosphere. These results validated measurements for wavelengths greater than 0.5 µm. More data on marine aerosol properties are needed before a quantitative determination of the ocean color spectrum from spacecraft altitudes can be made.

MICROWAVE SENSORS

The Skylab EREP instruments greatly expanded the microwave remote-sensing program from space by including both active and passive microwave instruments for measuring the scattering and emitting properties of the Earth's surface. The S193 instrument operated at a frequency of 13.9 GHz both as an altimeter and as a [316] combination radiometer/scatterometer. The latter instrument was designed to provide both vertical and horizontal polarization data at five incidence angles (0° to 48°). The S194 radiometer provided an additional frequency at 1.4 GHz for gross resolution studies of emission.

The altimeter sensor utilized the high-range-resolution capability of radar to measure the surface height variations along the satellite ground-track. The primary objective of the altimeter experiment was to establish its potential as a remote sensor of both the geoid over the ocean and the topography over land.

The backscatter experiment provided information to radar designers concerning the range of backscatter coefficients expected from space for a variety of surfaces, angles of incidence, polarizations, and geometric configurations. Also investigated was the dependence of the backscatter on the geometrical characteristics and the physical state of land areas and on the surface roughness over the oceans.

Similarly, it was of interest to determine the range of brightness temperatures that could be expected for different polarizations and incidence angles as sensed by a radiometer in space. Also to be investigated were relations between the brightness temperature and the physical state of land areas, the oceanographic surface parameters, and the atmospheric variations. Over ocean areas, the simultaneous RADSCAT observations could be used to provide corrections for the atmospheric attenuation of the backscatter measurements.

Several new techniques and special analysis procedures were developed to convert the raw microwave measurements to meaningful data for interpretation and comparison with other independent information. For the altimeter observations, special techniques were developed for (1) the accurate determination of the antenna pointing angle from the radar return, (2) the calibration and correction of the normalized radar cross section for pulsewidth/beamwidth-limited conditions, (3) deterministic and statistical analysis of the radar return for evaluating terrain reflection characteristics, and (4) the modeling of surface characteristics to reduce intrinsic noise of the altimeter height measurements and to separate the height biases.

Two approaches were used for the S193 RADSCAT data investigations. One was a general investigation in which brightness temperature and backscatter were categorized for the different incidence angles and polarizations to provide surface signature data for future microwave sensor design. In the second approach, the microwave signature was correlated with specific parameters such as ocean windspeed, soil moisture, and vegetation cover. Using the second method, investigators confronted a basic problem of determining the validity and accuracy of the ground-truth parameters, and several techniques were used to separate the uncertainty of the ground-truth parameter from the inferred parameter as measured by the microwave sensor from space (sec. 5).

The S194 radiometer data obtained over the oceans were used primarily to verify the theoretical relations of the microwave brightness temperature to ocean surface variables such as windspeed, sea-surface temperature, and salinity. The different techniques and analyses used to establish the relations between the measured microwave signature and the corresponding physical parameters of the Earth surface are discussed in the following subsections.

The S193 Altimeter Experiment

The altimeter precision needed for geodetic measurements over the oceans required a careful analysis of all potential error sources that contribute to the final error in the geodetic height determination. McGoogan et al. (ref. 6-55) derived the geodetic height h_g from

(6-17)

where h_s is the satellite height above a reference spheroid as obtained from the satellite tracking data; h_a is the altitude measured by the altimeter; and represents the dynamic ocean effects due to tides, winds, and currents. As indicated by McGoogan, can be considered negligible relative to the expected precision of approximately 1 m root mean square (rms), and no correction was made for it.

The major errors are due to orbital uncertainties and altimeter measurement inaccuracies. Because the altimeter is used primarily to determine the higher frequency geoidal components, methods for short-arc analysis were developed in which systematic error effects due to air drag, thrusting, and geopotential errors could be minimized. Comparison between short-arc analysis with extensive tracking coverage and longer arc....

....analysis with reduced tracking coverage showed that a significant bias and tilt could be introduced by orbit uncertainties (fig. 6-44). McGoogan developed techniques to estimate the orbital accuracy for each pass; and potential bias values from-20 to 135 m, depending on the quality and quantity of the tracking coverage, were obtained for different passes.

The error sources that contribute to the measured altimeter height h_a were separated into three main categories corresponding to the basic instrument delays, the pointing error, and the atmospheric path delay. The basic instrument errors were reduced by careful preflight and in-flight calibration of the system delays and by use of range-tracker design information.

To determine the pointing angle, two new techniques were developed. In one technique, the radar return was used directly for pointing-angle analysis. Because the altimeter operated in essentially a beam-width-limited condition, the mean radar return waveform consisted of a stretched pulse having a trailing edge that corresponded to the angular variation of the antenna gain. When the antenna pointed off-nadir, the increase of the antenna gain in the off-nadir direction increased the amplitude of the trailing edge relative to the peak amplitude. A plot of the theoretical relation for various off-nadir pointing angles is shown in figure 6-45(a), and a plot of a measured waveform is shown in figure 6-45(b). The pointing angle could be measured from the....

FIGURE 6-45.-Effect of pointing-angle variations on radar returns normalized to unity. (a) "Flat sea" return for no jitter. Height is 435 km (235 n. mi.). (b) Comparison of measured mean return and theoretical return for a 100-nanosecond transmitted pulse-width (pointing angle 0.5° off-nadir). Error bars are based on one standard deviation . [For a larger picture, click here]

---

....deviation of the trailing edge with an uncertainty of ±0.05°.

The waveform method can be used only when the off-nadir pointing angle is less than one-half a beam-width (0.6°); beyond this angle, the accuracy of the method is degraded because the leading edge starts to spread out. For larger off-nadir angles, the statistical characteristics of the range-tracker "jitter" (range fluctuations) were analyzed to deduce the pointing angle. Range jitter is produced by the intrinsic noise of the radar return due to the electromagnetic reflection properties of the ocean surface. An example of this technique is shown in figure 6-46. As mentioned previously, the leading edge stretches out as the off-nadir angle increases above 0.6°. The range-tracker-jitter amplitude and frequency response is related to the slope of the risetime, and increasing the risetime will increase the range-tracker jitter and reduce the bandwidth, as indicated in figure 6-46(c).

After the off-nadir pointing angle was determined, preflight calibration data were used to convert pointing angle offset to effective height corrections. The corrections and calibration of the data reduced the absolute rms height error from 20 to 10 m; however, relative errors for any given pass should be less than 1 m.

The corrected altimeter geoid height data were plotted for all operating passes and compared with a reference geoid (Goddard Earth Model 6 (GEM-6)) and with ocean-bottom topography where available. Several interesting relationships between the measured geoid and the sea trenches and mountains were discovered. (See sec. 5, entitled "Oceans and Atmosphere.")

Mourad et al. (ref. 6-56) used the calibrated altimeter data from four passes to further investigate the effects of the orbit and instrument noise on the results. A "best estimate" of the geoidal profile along the satellite groundtrack was obtained by applying the generalized least squares collocation method (GLSCM) to the...

....altimeter height observations. The analysis consisted basically of three steps.

1. The geoidal height was obtained as before (eq. (6-17)) by subtracting the measured altimeter height from the computed satellite height relative to a reference spheroid. This residual was then modeled as the sum of the expected random height signal and the intrinsic random range noise of the altimeter. For short arcs ( < 450 km), the signal can be expressed as the sum of a calibration constant reflecting the systematic orbit, altimeter and environmental errors, and the random height signal having variations bounded by the expected frequency characteristics of the global geoidal height distributions.

 

2. Using the signal spectral characteristics, the intrinsic range noise was reduced by applying the GLSCM to the residual values (filtering).

 

3. The filtered residuals were then compared with the Marsh-Vincent (M-V) geoid (Goddard geoid 73 (GG-73)), and by using weighting functions that corresponded to the reduced noise level of the altimeter measurements and the uncertainty of the M-V geoid, the GLSCM was reapplied to determine a calibration constant and the geoidal height deviations relative to the ground-truth M-V geoid. Several segments of the short-arc tracks were combined in the GLSCM solution by constraining the adjacent geoidal height values of the segments to be equal.

The values of the calibration constant reflecting potential altimeter and orbital biases varied between 20 and 50 m for passes 4 and 7 and compared with estimated orbital biases of 20 to 30 m (ref. 6-55) for the same passes.

Ground truth, consisting of the free-air gravity anomalies, the bottom topography, and the reference M-V geoid along the Skylab groundtrack, was assembled and compared with the derived geoid profile. An example of the results is shown in figure 6-47. The....

FIGURE 6-46.-Off-nadir pointing-angle effects on range and power spectrum. (a) Skylab 4 pass 38/86; = 0.6° (b) Skylab 2 pass 9; = 1.1° (c) Skylab 3 pass 21/32; = 1.4°. [For a larger picture, click here]

---

...relatively large and slow frequency fluctuations of the unfiltered altimeter geoid were caused by a large off nadir angle (1.25°; ref. 6-55). The application of the GLSCM effectively suppresses these noise fluctuations, as seen by the resultant filtered altimeter geoid. The depression of the altimeter geoid is displaced horizontally by approximately 200 km relative to the groundtruth geoid (GG-73) but agrees with the locations of the Puerto Rico Trench and the free-air gravity anomaly. Other results of Mourad's analysis that illustrate the altimeter sensitivity to the fine structure of the geoid are discussed in section 5.

The determination of values of the normalized radar cross section of the ocean at near-nadir angles and their accuracy was derived by Brown (ref. 6-57) from the altimeter data and calibration curves. The radar cross section is expressed by

(6-18)

where

is the incidence angle that produces the peak amplitude;

c is the speed of light

h is the altimeter altitude

L_P is the path loss

is the peak of the mean power as obtained from the automatic gain control (AGC) calibration curves at receiver temperature T_r

[322] is the in-flight received-power measurement obtained in the calibration data submode (CDS)

is the correction factor for the change of measured waveform relative to the waveform used in calibration, and the conversion from the mean of peak values of the received power measured by the altimeter to the peak of the mean values needed to compute

is basically the convolution of the system point-target response with the flat surface impulse response and is a function of the radar parameters, the antenna pattern, the pointing angle and the time delay at which maximum value is obtained

K represents the system losses obtained from preflight calibration

Brown (ref. 6-57) computed values of for each of the passes during which the pointing angles were less than 0.8°. The pointing angle was derived from the waveform return as described previously.

Using the pointing angle and the measured antenna pattern, the function could be computed. Because the antenna pattern was asymmetrical, two values for were computed, one with the pointing angle assumed to be in the roll directionand one assigned to the pitch direction . For Skylab 2 and 3, the values of as a function of are shown in figure 6-48. Typical values of for small incidence angles ( = 0.5°) varied between 8 and 16 dB.

Brown (ref. 6-57) also performed an error analysis to establish the uncertainty of the derived values of as a function of the pointing angle. Estimates of errors due to uncertainties in calibration, to , and to dominant bias were made, and the resultant error was computed. The Skylab 2 and 3 results show that absolute and relative rms errors of 0.7 and 0.3 dB, respectively, are achieved for a known pitch angle of 0° to 0.5°. When the antenna pattern was degraded, as in Skylab 4, rms errors of approximately 4 and 0.5 dB were obtained for absolute and relative values, respectively, with pitch angles of 0° to 0.5°.

Altimeter performance over terrain in the United States was investigated by Shapiro et al. (ref. 6-58) to evaluate the capability of the sensor to profile terrain....

---

FIGURE 6-48.- as a function of (0°, and ( ,0°) for Skylab 2 and 3.

---

...topography along the satellite groundtrack. The nonhomogeneity of terrain topography within a footprint reduced the potential height accuracy but provided relatively reliable tracking over most areas having small height variations. In general, the existence of high-reflectivity patches provided a waveform on which the split-gate range tracker could operate. The mean reflected power relative to the ocean return for different types of terrain is listed in table 6-V. The values decrease with increasing terrain complexity. The large received power at nadir over the salt flats is equivalent to a radar backscattering coefficient of 32 dB and, when compared with the Moore et al. (ref. 6-59) value of 14 dB at 1.5° incidence angle, indicates a highly specular return. Generally, a dominant specular return provided the required waveform for proper range-tracker operation as shown both by the dropoff of the backscatter between 0° (altimeter operation) and 1.5° (scatterometer operation) over the same type of terrain, and by the waveform analysis. In figure 6-49, a typical example of the radar return waveform from farmland in lowa is compared with the water return waveform from Lake Michigan. The terrain return is similar to a point-target return with a sharp trailing edge, whereas the trailing edge of the Lake Michigan return is stretched by the diffuse return of the water surface. Also shown is the correlation between the temporal (equivalent to a spatial resolution of 70 m) amplitude variations at the risetime portion of the individual radar returns (100/sec). The analysis shows that the amplitude of the water return decorrelated for a displacement of 70 m along the track, whereas the land return is correlated for....

.....several hundred meters. Thus, it can be deduced that specular patches of approximately a few hundred meters exist and that the altimeter measures the height to these bright spots within the altimeter footprint.

Figure 6-50 shows an altimeter output and topographic map profile. An example of the altimeter profile over Arizona is shown in figure 6-50(b) compared with the power return (fig. 6-50(a)) and an equivalent topographic profile (fig. 6-50(c)) obtained from a 1:250 000-scale contour map. The large received-power variations could not be correlated directly with the topographic variations because they also depend significantly on surface roughness, soil type, and soil moisture. Additional variations are caused by instrumental effects of the narrow AGC sampling gate.

The correlation between the altimeter profile and the map profile shown in figure 6-51(a) gives a correlation coefficient of 0.92 when the altimeter profile is shifted by approximately 7 km backwards relative to the map profile. A corresponding mean height difference of approximately-15 m and an rms height variation of ±35 m were computed, and the results are shown in figure 6-51(b). The backward shift has been detected in all passes and is probably due to the inertial response of the range tracker to height variations.

The results show that an altimeter, such as the S193, profiles the subsatellite groundtrack topography but acts as a low-pass filter (as opposed to altimeter geoidal operation over the ocean) because it responds primarily to the lower specular areas within the footprint. Future altimeters should provide for both specular and diffuse returns and should have additional sampling gates available so that the vertical structure within a footprint can be determined.

The S193 RADSCAT and S194 Radiometer Experiments

An overall evaluation of the RADSCAT measurements was made by Moore et al. (ref. 6-59) to determine the relationships of the measurement parameters to.....

.....polarization, incidence angle, soil moisture, type of terrain, and vegetation cover. Data obtained over the United States, Brazil, and the oceans were categorized, and specific test areas in which large anomalies were observed were used to establish potential correlation with the physical parameters of the observed areas.

The angular dependence (approximately 1.5° to 52° for the S193 radiometer) of the brightness temperature T_B over the continental United States is summarized in figure 6-52(a) for vertical-transmit/vertical-receive (VV) antenna polarization. Also shown is the number of samples for each data point and the upper and lower decile values. The mean brightness temperature over land is approximately constant at 268 K with a decile....

....value range of approximately ±15 K for all incidence angles. This relationship agrees with a Lambert's law model for very rough surfaces.

The mean backscatter response over land shown in figure 6-52(b) reveals a two-step dropoff with incidence angle; this relationship can be analytically expressed by a best fit to the data by

(6-19)

The decile values around the mean value are relatively small (4 dB) for angles larger than 10° because of spatial averaging of the large footprint (>100 km²).

The angular dependence of the brightness temperature over the oceans for both vertical and horizontal polarizations agrees with a slightly rough surface model, whereas the mean backscatter ocean response [326] follows a one-step dropoff given by

(6-20a)

for VV polarization and

(6-20b)

for horizontal-transmit/horizontal-receive (HH) polarization, where 0° < < 45°.

Statistical analysis of the different measurement parameters over land showed a low correlation between the brightness temperature and the backscatter coefficient and a high correlation between horizontal and vertical polarization as well as between the microwave backscatter measurements at different incident angles for > 15°. These results imply (1) that active and passive microwave measurements over terrain are sensitive to different surface and/or atmospheric characteristics, whereas the use of multiple polarization, at least for the given spatial resolution, is redundant, and (2) that side-looking radar performance can account for the far-range effects.

The backscatter coefficients are compared to land use categories for different incidence angles in figure 6-53. The values overlap for most land use categories, except for the high values over the salt flats in Utah, water surfaces at an incidence angle of 1.5°, and the low water values at incidence angles of 33° and 46°. An attempt to distinguish different types of terrain by a statistical decision procedure applied to the microwave data was unsuccessful for two reasons. The larger scatterometer footprint generally included different land use categories; and other factors, such as soil moisture, may dominate the microwave terrain response.

To establish a more precise correspondence between the physical state of the terrain and the microwave data, some uniform land areas over which large deviations in the microwave data were observed were studied in detail. A uniform rangeland groundtrack in Texas showed a large change in brightness temperature (288 to 236 K) and backscatter signal ( - 11 to -7 dB). To determine whether this change could be related to soil moisture, the pattern of precipitation along the groundtrack for 5 days preceding the overpass and on the day of the overpass was studied. Using the 5-day antecedent precipitation index (API) as an estimate of the soil moisture distribution along the satellite groundtrack, a correlation between the microwave data and the soil moisture was computed. The result indicated a relatively high correlation ( - 0.76 for land emission and 0.62 for backscatter). Results by Eagleman et al. (ref. 6-60) show that the correlation is improved if the API estimate of soil moisture is extended over 10 days.

Another area in which large changes in backscatter and emission were observed was the Great Salt Lake Desert in Utah. Brightness changes of approximately 70 K between the surrounding terrain and the desert were observed by the S193 sensor at a frequency of 13.9 GHz (fig. 6-54),by the S194 sensor at 1.4 GHz (fig. 6-55), and by the Nimbus-5 radiometer at 19.35 GHz. The Nimbus-5 spacecraft had passed over the area many times in 1972 and 1973. Corresponding large anomalies in backscatter are indicated in figure 6-53. To explain this large change in emission and backscatter over an apparently dry and smooth region, a two-layer surface model was assumed in which the low emission and the large backscatter could be associated with a subsurface brine layer having a dielectric constant considerably larger than that of the surrounding area. This model was based on the history of the region, which was originally covered by Lake Bonneville. The measurements indicated that the thickness of the dry surface layer may vary from 1 m to 10 cm over the Great Salt Lake Desert.

The large uniform forest and savanna-type areas in Brazil provided an opportunity to investigate the effect of different biomes on microwave radiation and scatter. The S193 backscatter pattern delineated the boundary between the relatively wet rain forests and the dryer savanna and thornbush region. A pseudoimage signature of the region obtained at VV and HH antenna polarizations is shown in figure 6-56. These detailed investigations have demonstrated that soil moisture is the most significant influence in terrain microwave emission and backscatter and that soil moisture may mask the roughness features of the observed terrain.

The S194 radiometer observations over water surfaces were evaluated by Hollinger-and Lerner (ref. 6-61) to determine the response of the radiometer to various oceanographic parameters. A rigorous and systematic method was developed to compute the expected antenna temperature and compare it with the measured values.

FIGURE 6-53.-Comparison of backscattering coefficient for various terrain categories at four incidence angles from Skylab 2 and 3 data obtained using the S193 scatterometer and VV antenna polarization. (a) = 1.5° (b) = 17° (c) = 33° (d) = 46°. [For a larger picture, click here]

---

The antenna temperature T_A was obtained from

(6-21)

where K is Boltzmann's constant, A_e is the effective antenna area, I_t, Q_t, U_t, and V_t are the Stokes parameters of the total radiation, l_a, Q_a, U_a, and V_a are the Stokes parameters of the antenna pattern in the proper reference frame, and is the solid angle over which the integration is performed.

The use of the Stokes parameters in equation (6-21) permits the cross-polarized components of the emitted, reflected, and atmospheric radiation to be simply summed and then to interact directly with the antenna polarization characteristics. A computer program was developed to compute the Stokes parameters as a function of the oceanographic and instrumental parameters in several steps.

First, the dielectric constant of seawater was computed as a function of radiometer frequency, sea-surface temperature, and salinity. From the computed dielectric values, the horizontal and vertical Fresnel reflection coefficients for water as a function of incidence angle were determined. The Fresnel reflection coefficients....

....were then modified to account for the surface roughness by using the empirical relation for differential brightness temperature

(6-22)

where U is the windspeed in knots and f is the frequency in gigahertz.

The Stokes parameters of the emitted and reflected components were then obtained by multiplying the proper combination of horizontal and vertical reflection coefficients by the equivalent black-body radiation. For....

.... the emitted component, the black-body radiation is obtained from the Rayleigh-Jeans approximation to Planck's black-body radiation law.

The sky radiation needed for the determination of the reflected component of the Stokes parameters was separately computed from simplified atmospheric models. For this purpose, the absorption coefficients for oxygen, water vapor, and liquid water were determined as a function of frequency and then integrated along the path of propagation to obtain both the downward-looking and the upward-looking radiation. The resultant values were then used to compute the reflected and atmospheric Stokes parameter components as well as the atmospheric attenuation. A final step in the program transformed the satellite coordinate system of the antenna pattern to the coordinate system of the observed surface so that the integration shown in equation (6-21) could be performed.

[331] The integration was made by mapping the S194 antenna beam pattern ( dB = 15°) by a grid of 1728 points out to 41° off-nadir so as to include the major side lobes. The height, the longitude, and the latitude at nadir were used as reference points for each observation; then, the position within the antenna beam was computed for every 10° in azimuth and 0.5° nadir angle to a maximum of 18°, and every 2° nadir angle to a maximum of 41°. For mixed land and water surfaces, each point was then identified as being on either land or water and the inputs to the models were adjusted accordingly. The land radiation was obtained from S194 measurements when land was known to fill the antenna beam completely. A separate program designed to compute the contribution of sunglint as a function of Sun elevation angle and windspeed was used primarily to eliminate extreme data points.

The S194 radiometer measurements were calibrated over three areas characterized by calm seas (windspeed U < 2.6 m/sec (5 knots)), minimum atmospheric loss, and availability of very good ground-truth data (passes 9 and 23). Using this calibration, the computed and measured antenna temperatures for all data (31 passes) were compared. The mean value of the difference is -0.0035 K with a standard deviation of 1.3 K. The small resultant error indicated that the theoretical relationships between the brightness temperature and salinity, sea-surface temperature, and windspeed could be directly used to predict the 21-cm radiometer sensitivity to these parameters.

The evaluation of the RADSCAT performance as a wind sensor over the ocean required the development of new techniques and models to provide the best estimate of the windspeed for a given measured value of the normalized radar cross section. Although previous theoretical and experimental investigations had indicated that the measured radar cross section is related to the sea-surface roughness and that the surface roughness in turn can be related to the surface windspeed, considerable scatter in the data precluded the establishment of quantitative relationships. The S193 RADSCAT observations provided, for the first time, a large data base, a wide range of wind conditions, and different instrumental and geometric configurations (polarization and incidence angles) so that the methods of statistical analysis could be effectively used.

The basic approach as used by Cardone et al. (ref. 6-62) was relatively simple. The values of the measured backscatter coefficients as a function of position and time were accumulated and compared with the corresponding surface-truth windspeed values. Then, using a theoretical relationship between backscatter and windspeed, the best-fit relationship between the two parameters was established. The implementation of this approach, however, was much more complex. In addition to the accumulation and evaluation of the large amount of data, the data had to be stratified according to quality and quantity, and according to the range of windspeed and windspeed direction, so that more meaningful and accurate relationships could be derived.

Because the quality of the conventional sea-surface-truth windspeed values was known to be highly variable, a special effort was made to improve the estimate of the wind vector at a given location and time by utilizing all available data sources as near to the location and time of the Skylab pass as possible. The wind data were classified according to quality (obtained from aircraft, weather ships, transient ships, etc.), and special models were used depending on the kind of weather systems and locations.

The analysis for Hurricane Ava illustrates the techniques used to acquire regional wind-field data. Beginning with the launch of Skylab, a systematic search was made for weather disturbances near the Skylab groundtrack to provide a large range of windspeeds for data interpretation. In early June 1973, it was apparent that a tropical storm would develop into a hurricane off the southwest coast of Mexico and that it would intersect with a Skylab pass on June 6. Preparations for obtaining S193 measurements were made, and the storm was tracked by U.S. Air Force aircraft. The National Oceanographic and Atmospheric Administration (NOAA) aircraft used in support of the Skylab observations was dispatched to Acapulco, Mexico, to By through the eye of the hurricane on June 6, the day of the expected Skylab pass. The measurements made by the aircraft (fig. 6-57) were then used directly as input boundary conditions to a hurricane model described by Cardone et al. (ref. 6-62) to derive a first estimate of the streamline-isotach distribution of the wind field. The modeled wind field was then refined by including all ship reports and data from other aircraft near the hurricane. The resultant composite analysis of the surface wind field is shown in figure 6-58. The solid lines (streamlines) are parallel to the wind direction, and the dashed lines (isotachs) are contours of constant.....

....windspeed. The circles indicate the position of the cells observed by the RADSCAT instrument. The wind vector at each position can be obtained directly from the chart by proper interpolation.

The backscatter coefficients were derived separately for each cell and time, corrected for atmospheric attenuation as obtained from the simultaneous radiometer measurements, and cataloged for each polarization and incidence angle. All measurements were referred to the five nominal incidence angles by correcting the values of the backscatter coefficients for small deviations from the nominal angles.

To establish the dependency of the backscatter coefficients on windspeed or of the windspeed on the backscatter coefficients, two methods were used. In the first method, all backscatter values were transformed to upwind direction by using the measured effects of wind direction (by the Advanced Applications Flight Experiment RADSCAT Program) as a conversion factor. Then, using the formal power relation obtained from an updated theory of backscatter, which included the effects of wind direction, a regression analysis was performed to minimize the variance between the predicted radar wind velocity U_r and the meteorological wind velocity U_m The relation used in the analysis was expressed in both logarithmic and power law form as

(6-23)

The computed coefficients b₀b₁ and for predicting the windspeed from the backscatter are shown in table 6-VI for the log model and in table 6-VII for the power law model, respectively. These tables show the sensitivity of calculations of the windspeed to changes in radar cross section as a function of both polarization and incidence angle, and indicate that the largest changes in backscatter for a given change in windspeed occur between incidence angles of 30° and 50°. The inverse of the parameters was also computed and showed that better agreement was obtained for the power law because of more uniform weighting of the lower windspeeds.

In the second method, the predicted radar wind is expressed as a combination of upwind, downwind, and crosswind velocity components and the aspect angle (the angle between the wind direction and the antenna direction). By an iterative method using both polynomial and power law relations, an effective power law is obtained for determining new coefficients, which are now functions of the aspect angle. The nonlinear relations used in the iteration process indicated that the deviations could be improved if the data were stratified for different wind-range intervals. The data were further stratified according to the source of the surface truth, such as weather research ships (type A) or aircraft, and transient ships (types B, C, and D), and were used for final error analysis.

To confirm the consistency of the results, the variance of the surface-truth values and the radar values of windspeed were first separately computed. The surface-truth windspeed variance was obtained using a withheld weather ship analysis, in which the value measured by the weather ship was assumed to be the true windspeed. The variance of the radar windspeed values was obtained by comparing the windspeed values obtained at the four different polarizations for each cell. It was then assumed that, if the results were consistent, the total variance of the difference between the meteorological wind and the radar wind should be the sum of the two variances. The results are shown in table 5-III in section 5. Residual variances are generally obtained only when the sample sizes are small. This finding increases the confidence that the assumed.....

[334] ....models and techniques used are valid and that the radar values, which have a constant variance of 0.98 m/sec (1.9 knots) squared, are generally more reliable than the values obtained from ship observations.

The brightness temperature and backscatter anomalies measured by the S193 over a groundtrack in Texas on June 5,1973, were related to the 5-day API by Moore et al. (ref. 6-59) as previously described. The 5-day API is a convenient moisture parameter because it requires only the daily precipitation, which is routinely reported by the weather stations. However, it is known that the actual soil moisture content may be considerably different from the value predicted by the 5-day API because it neglects the effects of the soil type and the runoff routing on the surface.

To establish a more accurate ground-truth base of soil moisture for both S193 and S194 measurements, Eagleman et al. (ref. 6-60) determined the soil moisture directly along the groundtrack of the satellite. Soil samples, collected at 7-km intervals near the time of the Skylab overflight for each 2.5-cm layer down to a depth of 15 cm, were weighed and dried in the laboratory, and the soil moisture (in percentage by weight) for each location and layer was determined. To compare the Skylab measurements with the ground-truth soil moisture values, the sampled soil moisture values were.....

[335] ....interpolated both in time and space by using the climatic water-balance technique developed by Thornthwaite and Mather (ref. 6-63). In this method, the potential evapotranspiration rate is estimated from the mean daily temperature and the value is adjusted to the day of the year and to the latitude. The actual evapotranspiration is then computed from the estimated potential evapotranspiration rate, the precipitation, and the available soil moisture, which depends on the soil type at the given location. Contour maps of the resultant soil moisture, such as that shown in figure 6-59(a), were then produced.

Similar distributions of S193 brightness temperature.....

....and backscatter were derived and are shown for the same test site in figures 6-59(b) and 6-59(c), respectively. The large footprint of the S194 radiometer observations did not permit presentation of such a high-resolution distribution, and a comparison between the S194 data and the ground-truth soil moisture was obtained by weighted averaging of the soil moisture content within the footprint.

The results of the analysis of five passes indicate that the highest correlation was obtained between the 21-cm (S194) brightness temperature and ground-truth soil moisture (fig. 6-60) because the 21-cm radiometer is more sensitive to soil moisture than the 2.2-cm RADSCAT and is less sensitive to surface roughness and atmospheric variations. The equivalent comparison between the 2.2-cm brightness temperature and soil moisture is shown in figure 6-61 for the Texas test site used by Moore (ref. 6-59). The correlation increases, in this case, from-0.76 with a 5-day API to-0.91 with the direct soil moisture content determination. A comparison of soil moisture with the backscatter measurements, however, shows a reduced correlation.

A comparison of the different sensor sensitivities to soil moisture is shown in figure 6-62 for the same size resolution cell. The poor response of the backscatter measurements is believed to be due to the large incidence angle (30°), at which roughness effects dominate the response. Other significant results of Eagleman's analysis indicate that horizontal polarization radiometry at 2.2 cm is less sensitive to soil moisture than is vertical polarization, that the best correlation is obtained with the top 2.5-cm-layer soil moisture, and that the height of the vegetation cover may modify the soil moisture measurements.

The performance of microwave sensors of soil moisture as a function of incidence angle was investigated by Stucky (ref. 6-21). For this purpose, the June 11 pass over Texas, made using sensors that operated in the in/rack-contiguous mode, was selected. The soil moisture parameter was expressed as the API for 11 and 6 days with a recession value (i.e., loss of moisture due to evapotranspiration arid subsurface runoff ) of 0.9. The daily precipitation was limited to 5 cm because it was assumed that any excess value would produce runoff and would not contribute to soil moisture. Interpolation between station API's and the API's at the footprint center was obtained by a distance-dependent circular weighting function that combined at least three station API's.

A relatively high correlation of soil moisture with 2.2-cm brightness temperature was obtained only for small incidence angles and for the 10-day API (fig. 6-63). The 6-day API neglected significant contribution to soil moisture of earlier precipitation, and at the larger angles, the effect of surface roughness and the atmospheric variations became more significant. Correlation coefficients between brightness temperature and backscatter decreased linearly with increasing incidence angle, starting with a maximum of-0.95 at 2° nadir angle.

McFarland (ref. 6-21) used the concurrent S194 measurements for comparisons with computed APl's.

Excellent agreement was obtained with the 11-day API as shown in figure 6-64, but another pass showed anomalies that could be correlated with irrigation and cultivation. Thus, although the 11-day API represents a good estimate of soil moisture, there are areas in which a remote sensor such as the S194 would produce a more accurate determination of the true soil moisture content.

---

---

---

[339] SUMMARY

The results of the Skylab EREP photographic experiment demonstrate that sensitometrically controlled, multiband photography constitutes a powerful tool for investigating Earth resources problems. The multiband characteristics of the photographic data were crucial for both visual and machine analysis of the data. The three most desired improvements in the photographic data are larger scales, better resolution, and more frequent coverage. The larger scales would make machine analysis of the data simpler because image densitometry of an object becomes easier as the image size of the object increases. Improved resolution would assist visual and machine interpretation and increase analysis accuracy. The desire for more frequent and repetitive coverage is not endemic to the photographic experiment; investigators using other sensors also expressed this need. However, the demonstrated ability to handle atmospheric and processing differences occurring between coverage dates has improved the potential value of multidate analyses and has undoubtedly led to increased requests for such coverage. Image digitization and computer analysis of photographic data offer perhaps the most significant and adaptable analysis approach for complex, multidate problems, and realization of the potential of this form of image analysis is only beginning.

The electrical recording of multichannel data from the S192 Multispectral Scanner enabled use of a wide range of signal-processing techniques and led to several developments in data manipulation by which the advantages inherent in multispectral remote sensing can be applied. In some cases, single bands outside the photographic region showed sufficient contrast and provided information for a particular use. Examples include the use of the 1.2- to 1.3-µm band for correlation to electrical conductivity or salinity, the thermal band to help differentiate commercial-industrial-residential land uses from natural vegetated areas, and the 1.55- to 1.75-µm band to separate waterfowl habitats from other land features.

The digital format enabled machine enhancement of detail by contrast-stretching a particular range of signals to match the display medium. By color translation and by overlaying images from two or more bands, improvements in discrimination of features in water, geological, and agricultural scenes were demonstrated.

The tape recording of the several channels also permitted the ratioing of a pair of channels that aided in the separation of ferric, ferrous, and nonferrous classes of materials. The ratio of spectral bands in the red/green region resulted in better correlation with suspended solids in reservoirs than either individual band. The ratio of infrared/red spectral bands was found useful in correlating soil moisture differences in the presence of partial vegetation cover.

The full potential of the multispectral data set was realized when computer analysis of the entire spectral range was performed. Subsets of optimum spectral bands were chosen by various statistical decision algorithms, and computer recognition of objects by using both supervised and unsupervised classification techniques was achieved with varying degrees of success. Modified clustering techniques led to reduced machine time, higher classification accuracies, and improved man/machine interactions.

Classification accuracies were improved by using different preprocessing rules. Atmospheric effects in the data were removed to improve accuracies, and atmospheric effects on the choice of channels were explored. Mixture-processing techniques were used to improve results whenever the resolution element contained a mixture of objects such as found near boundaries between classes. Signature-extension schemes were explored and information about elevation, aspect, and slope helped to improve recognition accuracies, particularly for scenes of mountainous regions. The use of statistical factor analysis in N-dimensional space, where N is the number of spectral bands, was explored and found valuable in the enhancement of rock outcrop and dense vegetation.

The S192 data also were used in studies of urban microclimate and in vertical wind profile analyses. The value of multiband scanner data was demonstrated especially for bands beyond those now available in Landsat. These should be incorporated in future space sensors.

The S191 infrared Spectrometer data were useful in obtaining information about the spectral transmission of the atmosphere and reflection characteristics of terrain classes. Ratios of spectral emissivities in the short and long-wavelength bands are useful in differentiating basaltic rocks from dacite. The S191 data were used in the study of the ocean color spectrum. Results agreed well with measurements for wavelengths greater than [340] 0.5 µm. More data on aerosol properties are needed to achieve similar results for wavelengths less than 0.5 µm.

The major objectives achieved by the EREP microwave investigators can be summarized in three areas: sensor performance evaluation, building of a data base for microwave sensor system design, and establishment of potential application areas.

The altimeter precision was validated, and methods for determining and correcting for pointing-angle and orbital errors were developed. Ocean radar cross sections for altimeter operation were determined with high precision, the electromagnetic reflection mechanism for both ocean and terrain was established, and measurements of terrain topography were shown to be feasible.

The scatterometer observations provided a large data base of backscatter coefficients as a function of surface reflecting and scattering properties, incidence angle, and polarization, and established the sensitivity of a scatterometer to windspeed for a large range of surface-truth windspeed values. Similarly, a large data base of brightness-temperature variations over the ocean and terrain was cataloged for different surface conditions, incidence angles, and polarizations. Theoretical relationships between variations of the physical ocean parameters of salinity, surface wind, and sea-surface temperature and the ocean brightness temperature were developed and verified with the S194 measurements. Potential application for determining soil moisture by long-wavelength radiometer observations was confirmed.

The results of the microwave investigations have both short- and long-term implications. The experience gained with the Skylab altimeter was immediately applied to the design and operation of the Geodetic Earth Orbiting Satellite C altimeter, which is now in orbit. Further refinements are planned for the Seasat altimeter (to be launched in 1978). The Seasat spacecraft will also include an improved version of a scatterometer that can determine both windspeed and wind direction.

The results of the Skylab altimeter terrain observations provide basic information for improved surface topography determination by altimetry if the altimeter sampling capability of the radar return is expanded. This concept is being considered for the Space Shuttle, which will orbit the Earth, and for unmanned spacecraft that will orbit the Moon and the planets.

The catalog of backscatter coefficients and brightness temperatures will help radar and radiometer designers to provide optimum microwave system performance for a variety of applications.

Finally, the availability of higher spatial resolution performance of higher spatial resolution performance of passive microwave sensors at the longer wavelengths will enable significant global synoptic measurements of soil moisture content.

REFERENCES

6-1. Potter, A. E.; Grandfield, A. L.; and Williams, C. K.: Flight Performance of the Skylab Earth Resources Experiment Package. The Skylab Results, Advances in the Astronautical Sciences, vol. 31, pt. 2, William C. Schneider and Thomas E. Hanes, eds., American Astronautical Society (Tarzana, Calif.), 1975, pp. 605-633.

6-2. Pirie, Douglas M.; and Steller, David D.: California Coastal Processes Study. NASA CR-144489, 1975.

6-3. Barnes, James C:; Smallwood, Michael D.; and Cogan, James L.: Study to Develop Improved Spacecraft Snow Survey Methods Using Skylab/EREP Data. NASA CR-144388, 1975.

64. Szekielda, Karl-Heinz: Dynamics of Plankton Populations in Upwelling Areas. NASA CR-144480, 1975.

6-5. Piech, Kenneth R.; Schott, John R.; and Stewart, Kenton M.: S190 Interpretation Techniques Development and Applications to New York State Water Resources. NASA CR-144499, 1975.

6-6. Baldridge, Paul E.; Goesling, P. H.; et al.: Utilizing Skylab Data in On-Going Resources Management Programs in the State of Ohio. NASA CR-134938, 1975.

6-7. Yarger, Harold L.; and McCauley, James R.: Skylab Study of Water Quality. NASA CR-144505, 1975.

6-8. Gordon, Hayden H.; and Nichols, Maynard M.: Skylab MSS Versus Photography for Estuarine Water Color Classification. Water Color and Circulation Southern Chesapeake Bay, Part 11. NASA CR-141404, 1975.

6-9. Welby, Charles W.; and Lammi, J. O.: Utilization of EREP Data in Geological Evaluation, Regional Planning, Forest Management, and Water Management in North Carolina. NASA CR-144104, 1975.

6-10. Hardy, Ernest E.; Skaley, J. E.; et al.: Enhancement and Evaluation of Skylab Photography for Potential Land Use Inventories. Part I and Part II. NASA CR-144473, 1975.

[341] 6-11. Colwell, Robert N.; Benson, Andres S.; et al.: Agricultural Interpretation Technique Development. NASA CR-144486, 1975.

6-12. Silva, Leroy F.: A Study of the Utilization of EREP Data From the Wabash River Basin. NASA CR-147412, 1976.

6-13. Gumerman, George J.; Hanson, John A.; et al.: The Hydrology of Prehistoric Farming Systems in a Central Arizona Ecotone. NASA CR-144492, 1975.

6-14. Plogg, F.; and Garrett, C.: Explaining Variability in Prehistoric Southwestern Water Control Systems. Paper presented at the 35th Annual Meeting of the Society for American Archeology (Mexico City, Mexico), 1970.

6-15. Langley, Philip G.; and Van Roessel, Jan: The Usefulness of Skylab/EREP S-190 and S-192 Imagery in Multistage Forest Surveys. NASA CR-147439, 1976.

6-16. Maul, George A.; Gordon, Howard R.; et al.: An Experiment to Evaluate Skylab Earth Resources Sensors for Detection of the Gulf Stream. NASA CR-147454, 1976.

6-17. Gilmer, David S.; and Work, Edgar A., Jr.: Utilization of Skylab (EREP) System for Appraising Changes in Continental Migratory Bird Habitat. NASA CR-147542, 1975.

6-18. Vincent, R.; Wagner, H.; Pillars, W.; and Bennett, C.: Assessments of Mapping Exposed Ferrous and Ferric Iron Compounds Using Skylab EREP Data. NASA CR-144504, 1976.

6-19. Vincent, R. K.; and Pillars, W. W.: Skylab S-192 Ratio Codes of Soil, Mineral, and Rock Spectra for Ratio Image Selection and Interpretation. Proceedings of the Ninth International Symposium on Remote Sensing of Environment, Environmental Research Institute of Mich. (Ann Arbor, Mich.), 1974, pp. 875-896.

6-20. Curran, Robert J.; Salomonson, Vincent V.; and Shenk, William: The Application of Satellite Data in the Determination of Ocean Temperatures and Cloud Characteristics and Statistics. NASA CR-147539, 1976.

6-21. Pitts, David E.; Sasaki, Yoshikazu; and Lee, Jean T., compilers: Severe Storm Environments: A Skylab EREP Final Report. NASA TM X-58184, 1976.

6-22. Phillips, T. L., ed.: LARSYS Version 3.1 User's Manual. Vol. 2. Laboratory for Applications of Remote Sensing, Purdue Univ. (West Lafayette, Ind.), 1973.

6-23. Swain, P. H.: Pattern Recognition: A Basis for Remote Sensing Data Analysis. NASA CR-130757, 1973.

6-24. Goetz, A. F. H.; Abrams, M. J.; et al.: Comparisons of Skylab and Landsat Images for Geologic Mapping in Northern Arizona. NASA CR-147503, 1976.

6-25. Sattinger, 1. J.; Sadowski, F. G.; and Roller, N. E. G.: Analysis of Recreational Land Using Skylab Data. NASA CR-144471, 1976.

6-26. Nalepka, R. F.; Morganstern, J.; et al.: S-192 Analysis: Conventional and Special Data Processing Techniques. NASA CR-144506, 1975.

6-27. Simonett, David S.: Application of Skylab EREP Data for Land Use Management. NASA CR-147457, 1976.

6-28. Eberlein, R.; Vanderburg, G. L.; Rosenfeld, A.; and David, L. S.: Edge and Line Detection in ERTS Imagery: A Comparative Study. NASA CR-138963, 1974.

6-29. Anderson, Theodore W.: An Introduction to Multivariate Statistical Analysis. John Wiley & Sons, Inc., 1958.

6-30. Rao, C. R.: Linear Statistical Interference and Its Applications. John Wiley & Sons, Inc., 1965.

6-31. Swain, P. H.; Robertson, T. V.; and Wacker, A. G.: Comparison of the Divergence and B-Distance in Feature Selection. LARS Information Note 020871, Purdue Univ. (West Lafayette, Ind.), 1971.

6-32. Hoffer, Roger M.: Computer-Aided Analysis of Skylab Multispectral Scanner Data in Mountainous Terrain for Land Use, Forestry, Water Resource, and Geologic Applications. NASA CR-147473, 1975.

6-33. Thomson, F. J.: Machine Processing of S-192 and Supporting Aircraft Data: Studies of Atmospheric Effects, Agricultural Classifications, and Land Resource Mapping. NASA CR-144503, 1975.

6-34. Wiegand, Craig L.; Richardson, Arthur J.; et al.: Soil Salinity Detection. NASA CR-144403, 1975.

6-35. Hannah, John W.; Thomas, Garland L.; Esparza, Fernando; and Millard, J. J.: Planning Applications in East Central Florida. NASA CR-145415, 1975.

6-36. McMurtry, George J.; and Petersen, Gary W.: Interdisciplinary Applications and Interpretations of EREP Data Within the Susquehanna River Basin. NASA CR-147541, 1976.

6-37. Selzer, Robert H.: The Use of Computers to Improve Biomedical Image Quality. Proceedings of Fall Joint Computer Conference, vol. 33, pt. 1, Dec. 1968, pp. 817-834.

6-38. Horwitz, Harold M.; Nalepka, Richard F.; Hyde, Peter D.; and Morganstern, James P.: Estimating the Proportions of Objects Within a Single Resolution Element of a Multispectral Scanner. Proceedings of the Seventh International Symposium on Remote Sensing of Environment, Environmental Research Institute of Mich. (Ann Arbor, Mich.), 1971, pp. 1307-1320.

[342] 6-39. Poulton, Charles E.; and Welch, Robin I.: Plan for the Uniform Mapping of Earth Resources and Environmental Complexes From Skylab Imagery. NASA CR-144484, 1975.

6-40. Houston, R. S.; Marrs, R. W.; and Borgman, L. E.: Multidisciplinary Study of Wyoming Test Sites. NASA CR-147719, 1975.

6-41. Haefner, Harold: Snow Survey and Vegetation Growth in the Swiss Alps. NASA CR-147395, 1976.

6-42. Goldman, Gary C.; and Horvath, Robert: Oil Pollution Detection and Monitoring From Space Using Skylab. NASA CR-144502, 1975.

6-43. Trumbull, James V. A.: The Utility of Skylab Photointerpreted Earth Resources Data in Studies of Marine Geology and Coastal Processes in Puerto Rico and the Virgin Islands. NASA CR-147437, 1975.

6-44. Polcyn, Fabian C.; and Lyzenga, David R.: Skylab Remote Bathymetry Experiment. NASA CR-144482, 1976.

6-45. Pease, Robert W.: Mapping Terrestrial Radiation Emission With a Scanning Radiometer. Proceedings of the Seventh International Symposium on Remote Sensing of Environment, Environmental Research Institute of Mich. (Ann Arbor, Mich.), 1971, pp. 501-510.

6-46. Villevieille, A.; and Weiller, A. B.: The Possibility of Evaluating Vertical Wind Profiles From Satellite Data. NASA CR-147475, 1975.

6-47. Aldrich, Robert C.; Dana, Robert W.; et al.: Evaluation of Skylab (EREP) Data for Forest and Rangeland Surveys. NASA CR-147440, 1975.

6-48. Polcyn, Fabian C.; Rebel, Diana L.; and Colwell, John E.: Analysis of Hydrological Features of Portions of the Lake Ontario Basin Using Skylab and Aircraft Data. NASA CR-147456, 1976.

6-49. Colwell, J. E.: Bidirectional Spectral Reflectance of Grass Canopies for Determination of Above Ground Standing Biomass. Ph. D. Dissertation, Univ. of Mich. (Ann Arbor, Mich.), 1973.

6-50. Savastano, K. J.: Application of Remote Sensing for Fishery Resources Assessment and Monitoring. NASA CR-147507, 1975.

6-51. Anding, David C.; and Walker, John P.: Use of Skylab EREP Data in a Sea-Surface Temperature Experiment. NASA CR-144479, 1975.

6-52. Malila, William A.; and Nalepka, Richard F.: Atmospheric Effects in ERTS-1 Data, and Advanced Information Extraction Techniques. Proceedings of the Symposium on Significant Results Obtained From the Earth Resources Technology Satellite-1, Goddard Space Flight Center (New Carrollton, Md.), Mar. 5-9, 1973, vol. I, sec. B, pp. 1097-1104.

6-53. Tingey, David L.; and Potter, John: Quantitative Determination of Stratospheric Aerosol Characteristics. NASA CR-147444, 1975.

6-54. Anding, D. C.; Kauth, R.; and Turner, R. E.: Atmospheric Effects on Infrared Multispectral Sensing of Sea-Surface Temperature From Space. NASA CR-1858, 1971.

6-55. McGoogan, J. T.; Leitao, C. D.; and Wells, W. T.: Summary of Skylab S-193 Altimeter Altitude Results. NASA TM X-69355, 1975.

6-56. Mourad, A. G.; Gopalapillai, S.; Kuhner, M.; and Fubara, D. M.: The Application of Skylab Altimetry to Marine Geoid Determination. NASA CR-144372, 1975.

6-57. Brown, G. S.: Reduced Backscattering Cross Section () Data From the Skylab S-193 Radar Altimeter. NASA CR-141401, 1975.

6-58. Shapiro, Allan; Thormodsgard, June M.; and Okada, J. M.: Skylab Altimeter Observations Over Terrain. NASA CR-144498, 1975.

6-59. Moore, Richard K.; Ulaby, Fawwaz T.; et al.: Design Data Collection With Skylab Microwave Radiometer-Scatterometer S-193. Vols. I and II. NASA CR-144537 and NASA CR-144538, 1975.

6-60. Eagleman, J. R.; Lin, W.; et al.: Detection of Soil Moisture and Snow Characteristic From Skylab. NASA CR-144485,1975.

6-61. Hollinger, James P.; and Lerner, Robert M.: Analysis of Microwave Radiometric Measurements From Skylab. NASA CR-147442, 1975.

6-62. Cardone, Vincent J.; Young, James D.; et al.: The Measurement of the Winds Near the Ocean Surface With a Radiometer-Scatterometer on Skylab. NASA CR-147487,1976.

6-63. Thornthwaite, C. W.; and Mather, C. W.: Instructions and Tables for Computing Potential Evapotranspiration and Water Balance. Publications in Climatology, vol. 10, no. 3, 1957, pp. 185-311.

² Robert H. Alexander, John E Lewis, Jr., et al, "Applications of Skylab Data to Land use and Climatological Analysis," unpublished Final Report, NASA-USGS Agreement T-5290-B, 1976.

---