Viewing Geometry of AVHRR Image Composites Derived Using Multiple Criteria

The U.S. Geological Survey currently generates composites of AVHRR imagery based on a single objective-maximizing the Normalized Difference Vegetation Index (NDVI)-as a means of reducing cloud contamination. Our research supports the findings of others that, i n some cases, NDVI is maximized at the expense of optimal viewing geometry; that is, satellite zenith angles are often further off-nadir than necessary to ensure cloud-free viewing. We explore various compositing methods by systematically varying weights on NDVI, satellite zenith angle, and maximum apparent temperature. A test composite of California from September 1990 appears to be superior to the maximum NDVI and maximum apparent temperature composites in several respects. First, the satellite zenith angle distribution is more closely clustered about nadir, which minimizes atmospheric path length, spatial distortion, and bidirectional reflectance effects. Second, neighboring pixels are more frequently selected with similar viewing geometry and atmospheric conditions. Introduction Compositing is a strategy for removing cloud contamination and atmospheric effects from a series of images over a discrete period of time. The U.S. Geological Survey's (USGS) EROS Data Center (EDC) has an operational program for developing 1-km resolution Advanced Very High Resolution Radiometer (AVHRR) data sets over the conterminous U.S. (Eidenshink, 1992). The conterminous AVHRR time series data are produced using the maximum value compositing (MVC) technique based on the Normalized Difference Vegetation Index (NDVI) (Holben, 1986). The highest vegetation index value derived from daily georeferenced images is selected for each pixel location in a 14-day period. Clouds and their shadows have low NDVI values relative to the true index value of vegetated surfaces. In theory, the maximum value technique leads to the selection of the greenest and, therefore, least cloud-contaminated data value. Based on this method, a biweekly composite data set is created that includes the five satellite data channels, the vegetation index, viewing geometry, and date of pixel observation. This time series of AVHRR data is distributed by EDC on CD-ROM for the years 1989 through 1993 (Eidenshink, 1992). NDVI values, which range from -1 to $1, are rescaled to 8-bit data between 0 and 200 in the EDc time series data. The compositing strategy currently being applied to create a global 1-km dataset, commencing with 1992 imagery, is similar except that the compositing period is only 10 days and a constraint to eliminate satellite zenith angles greater than 42" was added (Eidenshink and Faundekn, 1994). When the MvC algorithm was originally proposed, it was argued that the method would preferentially select near-naDepartment of Geography and Institute for Computational Earth System Science, University of California, Santa Barbara, CA 93106-4060 ([email protected]). PE&RS June 1997 dir views over larger scan angles (Holben, 1986). The work that led to that recommendation was based on a model using an assumption of Lambertian reflectance from Earth surfaces. Numerous studies have found that off-nadir viewing can produce greater NDVI values than at nadir viewing angles. Deering and Eck (1987) found the highest NDVI in the forwardscatter direction for grass and soybean cover, in large part caused by greater shadowing from the plant canopy in this direction. The scan direction of the afternoon satellites used for compositing also is near the principal plane of the sun in the northern hemis~here. which can accentuate the amount of shadowing from i h e canopy even further (Goward et al., 1991). Yang et al. (1996) observed a similar forward-scatter bias in the North American portion of the global 1-km dataset averaged over one year, but not for the South American composites. Possibly, the relationship with the principal plane of the sun is responsible for this geographic difference. Gutman (1991) showed strong forward-scatter bias in a 10day NDVI composite over the Great Plains using a 40-km resolution dataset. Cihlar et al. (1994) observed that the MVC algorithm preferentially selected forescattering views over nadir for a variety of cover types in central Canada. They attributed this to a higher transmittance of near infrared radiation through the canopy compared with red radiation. On the east coast of the United States, results have been contrary (Allen et al., 1994; Moody and Strahler, 1994). Satellite zenith angle distribution was biased towards backscatter views. Moody and Strahler (1994) suggested that this bias was caused by two factors: that westward-looking or forward-scatter views may not have been available in the dataset and that scanning in New England was not near the principal plane of the sun. The first factor was presumably a result of the AVHRR data being acquired at the EDC in Sioux Falls, South Dakota, and consequently limiting the number of dates, and angles, available. The relationship of viewing to the principal plane of the sun would reduce the main anisotropic effects from the atmosphere and the surface. Certainly the combination of off-nadir bias in compositing and regional variations in effects would make it more difficult to interpret NDVI derived from AVHRR composites over the conterminous U. S. or globally. Biophysical modelers would encounter an added level of uncertainty when using these composites as inputs. Off-nadir viewing causes several potential problems for interpreting NDw. For off-nadir views, the Instantaneous Field of View (IFOV) integrates surface leaving radiance from a larger footprint on the ground than the output pixel area, and it overlaps the IFOV of neighboring pixels. Furthermore, the scattering properties of the atmosphere are directional, Photogrammetric Engineering & Remote Sensing, Vol. 63, No. 6, June 1997, pp. 681-689. 0099-1112/97/6306-681$3.00/0 O 1997 American Society for Photogrammetry and Remote Sensing meaning that views towards the sun will be affected by quite different scattering than views away from the sun at the same viewing angle. Most Earth surface materials also exhibit an anisotropic property, further distorting the reflectance observed at the satellite from the true values at ground level. Therefore, it is important to more fully understand the characteristics of the NDVI time series data from EDC with respect to viewing geometry. Concurrently, the need to explore alternative compositing strategies has been widely recognized (Viovy et al., 1992; Cihlar et al., 1994; Eidenshink and Faundeen, 1994; Townshend et al., 1994). We undertook an analysis of the 1990 AVHRR time series to examine its viewing characteristics on the west coast of the United States. We also experimented with alternative compositing algorithms to reduce the potential satellite zenith angle bias. The objectives of this paper are (1) to quantify the viewing geometry of the 1990 composites derived by the maximum NDVI value algorithm and interpret these findings in relation to land cover for California, and (2) to evaluate alternative algorithms that systematically vary the importance of three criteria in the compositing process. The distributions of satellite zenith angles from the alternatives are compared. Three alternatives are compared: MVC, maximum apparent temperature ( M ~ T ) (Cihlar et al., 1994), and one that includes both maximum apparent temperature and satellite zenith angle criteria. We refer to this multiple objective compositing algorithm by the acronym MOC. Background The EDC NDvI time series has been used for a wide variety of applications, including land-cover mapping and characterization (Loveland et al., 1991; Kremer and Running, 1993; Paruelo and Lauenroth, 1995), crop assessments (Wade et ul., 1994), fire monitoring (Kasischke et al., 1993), biodiversity assessment (Walker et al., 1992), and interannual variation in plant phenology (Reed et al., 1994). For many if not all of these applications at regional and larger scales, it is essential that data be compiled consistently so that NDVI has a constant relationship with biophysical parameters. The AVHRR sensor achieves its daily global coverage by means of scan sweeps to over 55 degrees off-nadir, producing a look angle of 68 degrees relative to the Earth's surface. Off-nadir viewing causes at least three potential problems for computing and interpreting NDVI: variations in spatial resolution, variations in atmospheric attenuation associated with changes in path length, and variations in bidirectional reflectance effects (Goward et al., 1991). Whereas a pixel at nadir has a spatial resolution of 1.1 km, at the extreme angle the effective resolution is 2.4 km by 6.5 krn. Adjacent pixels of off-nadir observations also have significant overlap. Goward et al. (1991) calculate that the ~ F O V becomes significarilly distorted beyond off-nadir zenith angles of 25 degrees. The information for each 1.21-kmz pixel in the geometrically corrected output is derived from an actual area of up to 16 km2 on the ground. All satellite remote sensing is subject to the atmospheric attenuation of surface-leaving radiance. Off-nadir viewing increases this effect by lengthening the path that the reflected energy traverses. Furthermore, the scattering properties of the atmosphere are directional, meaning that views towards the sun will be affected differently than views away from the sun at the same angle. Aerosol scattering is more pronounced in the red channel of AVHRR (Channel 1) than in the near infrared (Channel 2), and is strongest in the backscatter direction (eastward looking on the afternoon satellite passes) (Deering and Eck, 1987; Holben et al., 1986). Most Earth-surface materials also exhibit anisotropic reflectance characteristics, further modulating the reflectance observed at the satellite relative to the true value at ground level. Holben et al. (1986), for instance, showed that NDVI of grasses tends to be lowest at extreme scan angles in the backscatter direction and highest in the forward scatter direction. The combined effect then, for equivalent atmospheric and surface conditions, is that NDVI will tend to be lower in the backscatter direction, and will be highest in the forescattering direction, or westward looking, than in either the eastward or nadir views (Goward et al., 1991). The potential impacts of viewing angle on NDVI and on cover-type classification led us to review the remote sensing literature for alternative compositing algorithms that might guarantee close-to-nadir composites. In a comparison of five algorithms, Cihlar et al. (1994) found that maximum apparent temperature produced composites which were consistently closest to nadir for crops and relatively close for forest types in Canada. Also, neighboring pixels were more often selected from the same scene, reducing artificial texture in the NDVI composite. Equally important, maximum apparent temperature produced NDVI values closest to a reference image. Another advantage of this method is that it uses data within the dataset rather than depending on ancillary meteorological or environmental data which may be incomplete. The maximum-apparent-temperature algorithm is based on the observation that clouds decrease the apparent surface temperature. Off-nadir pixels would also tend to be excluded if other choices were available because the increased path length further attenuates the emitted thermal signal. It could be subject to rejection of nadir views if surface conditions affect apparent temperature more than the atmospheric effects do, e.g., rainfall over bare soil. The algorithm used by Cihlar et a]. (1994) is based on AVHRR Channel 4, in the 10.3 to 11.3 micrometre range. As a second choice, they also recommend a two-step procedure using the maximum NDvI followed by a minimum scan angle criterion applied to all pixels within 15 percent of maximum NDVI in that period. Viewing Geometry of the 1990 Time Series Composites For the state of California, the satellite zenith angle files for the 19 composites of the 1990 time series were obtained from the CD-ROM (USGS, 1991) and averaged for every pixel. From visual examination of this derived image, there appeared to be a probable relationship between viewing geometry of the NDVI data and both vegetation and topographic features. In order to quantify these apparent relationships, the CALVEG map of plant series (Matyas and Parker, 1980) was recoded to general types, i.e., subformations, and gridded to the same resolution as the A V H R R ~ ~ S ~ ~ data. An identical 10 percent random sample was taken from the NDVI images and the vegetation map for statistical analysis in S-PLUS (Statistical Sciences, Inc., 1991). The size of the sample (greater than 40,000 samples) was large enough that all types would be proportionally represented. Therefore, stratifying the sampling by vegetation types was deemed unnecessary. Data were analyzed by examining the frequency histograms of NDvI and viewing geometry and by creating boxplots of angles by land-cover class. Alternative Cornpositing Strategies To evaluate alternative compositing algorithms, we obtained the daily georectified AVHRR data from EDC. The California study area was windowed from reflectance Bands 1 and 2, thermal Band 4, and the satellite zenith angle files for each daily image. The composite period of 14 September to 27 September 1990 was selected for several reasons. It is late in the summer dry season in California's Mediterranean climate, and the contrast between many vegetation types is high then. The landscape is not saturated with peak green-