Mapping land-use has been an important application of remote sensing and has generally been accomplished using traditional classification techniques ( i. e., supervised/unsupervised classifications.) When properly applied, these methods have been successful with individual image data sets. However, because of differential atmospheric, illumination, and instrumental effects, it is difficult to obtain consistent classes with these approaches between images taken at different times.
Another approach to classifying/monitoring land-use using multispectral images is based upon spectral mixture analysis (SMA). In this method, radiance is transformed into fractions of spectral endmembers (reflectance spectra of known materials in the image). Interpretation of these endmember-fractions (i. e. , green vegetation, soil, non-photosynthetic vegetation, shade) is more intuitive than radiance and lends itself to a physically-based image interpretation. These fractions can be used as a framework for classification and/or monitoring surface changes over time.
One application of SMA is for image classification where classes are defined as a domain of fractions that identify a type of land-use. An example of this application is a study to monitor land-use from 1988 through 1991 in an area north of Manaus, Brazil at two cattle ranches, Fazenda Dimona and Fazenda Esteio (Adams et al., 1995). These areas are dominated by primary forest that have been cleared at different times to create pasture. As a result, these areas contain large fields in various stages of regrowth.
Methods
A subset of four Thematic Mapper (TM) image data sets for this study area were used (1988, 1989, 1990, and 1991.) All four images were taken in August, the most cloud-free time of the year. The images were co-registered and then calibrated by: 1) calibrating the 1989 image using the spectral mixture analysis method (Smith et al., 1990), and then 2) intercalibrating the remaining images to the 1989 calibrated image. Spectral mixture analysis was applied to all four images using four reference endmember spectra that represented the 4 major surface components of natural environments: 1) green vegetation, 2) non-photosynthetic vegetation (i. e., bark, senescent grass, litter, twigs), 3) soil, and 4) shade (Gillespie et al., 1990.)
The data cluster of fractions for the 1989 data set were used to define class boundaries us ing an interactive computer display. The cluster of fractions were displayed on a modified tethedron that could be rotated to any projection. Each data point in the tethedron corresponded to a pixel in the image. Pixels highlighted in either the tethedron plot or image display, were simultaneously highlighted in the other. In this way, the limits of different classes were initially identified by using areas of known land-use from field observations. The boundaries were then refined by testing the classification on the three remaining images.
Once all the images were classified, a pixel-by-pixel comparison of class changes over time was performed. In a few cases, unlikely class changes were observed. For example, it is highly improbable that an area of bare soil one year would become mature ( terra firme) forest the following year. For these pixels, the class history was examined. An aberrant classification in a single year of a physically understandable regrowth trend would be reevaluated and reclassified. Tracking the pixel history was an effective tool for improving image classification.
Once all the images are classified, pixels with specific histories can be identified. For example, one can identify all of the pixels in an image that, at one time, were forest or regrowth and were subsequently cut.
Results
The simple mixing model used in this study accounted for 95% + of the spectral variability in the four image data sets. However, because only 4 endmembers were used to model the image, (with only single representative green vegetation [GV] and non-photosynthesizing vegetation [NPV] spectra), spectral mimicking made certain classes inseparable. For example, dry pasture and slash, which both have high fractions of NPV, were spectrally indistinguishable. Therefore, to minimize confusion caused by applying inappropriate class names, the term "category" was used. Each category included one or more spectrally similar classes. In this study, 8 categories were identified:
Many of the pixels in the image were assigned unambiguous class names by monitoring the changes of the categories over time for each image pixel or by image context ( i. e., long pixel-wide areas of category 7 [bare soil /roads], would be classified as a road.) Field work showed that the classification accuracy was consistently high.
In the study area, a trend of overall regrowth was identified. Many of the areas that were cut for pasture were subsequently allowed to evolve back to a secondary forest. Some pastures had been maintained through subsequent cutting/burning. Although the method was tested in the Amazon basin, the results suggest that endmember classification may be generally useful for comparing multispectral images in space and time.
The Jasper Ridge, California Example
The same approach to image classification was performed on the 2 June 1992 Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) (224 spectral bands) image taken over the Jasper Ridge Biological Reserve near Palo Alto, California. This area, with a Mediterranean climate, is covered with a variety of forest canopies and grasslands, including deciduous oak woodland, chaparral, evergreen forest, and forested wetland. The image was calibrated to reflectance using Green et al. (1993) and modeled using an appropriate set endmember spectra. Again, four endmembers were used representing the same four major surface components: green vegetation, NPV, soil, and shade. The same classification scheme was used as with the Brazil example, utilizing the same classification boundaries. The category names were changed to the appropriate land-cover names for this study area. A comparison of the resulting classified image to vegetation maps and subsequent field investigations showed a very high degree of classification accuracy. Although the method was tested in the Amazon basin, the results suggest that endmember classification may be generally useful for comparing multispectral images in space and time.
Spectral Trajectories
Imbedded in the change in classification over time is the change in fractions. The temporal change in fractions can be useful in monitoring surface processes. Seasonal changes of forest canopies and grasslands have been detected in Jasper Ridge, California (Sabol et al., 1993) between June and October 1992. An increase in NPV at the expense of green vegetation from June through October was due to increasing exposure of bark and stems as deciduous trees drop their leaves. By understanding the seasonal spectral changes, one may be able to separate and identify subtle long-term ecological changes in forested areas.
The same approach has been used to map the status of forest regrowth in cut areas of the Gifford Pinchot National Forest in Washington and to establish fractional regrowth trends (Sabol et al., 1995). Recent clearcuts were characterized by high fractions of NPV and low fractions of GV and shade. With in creasing cover, stands had correspondingly higher fractions of GV and shade, along with a decrease in the fraction of NPV. Closed-canopy stands had low fractions of NPV (contributed by exposed branches) and intermediate values of shade. Older stands with increasingly complex canopies had higher fractions of shade.
Conclusion
Spectral mixture analysis is an intuitive, more interpretable framework for image analysis. Natural environments can be well modeled using the simple mixing model (with green vegetation, NPV, soil, and shade endmembers). The resulting fractions can then be used to identify (classify) different land-uses and/or regrowth states. When applied over a time series, surface processes can be monitored. This approach is independent of imaging system and is generally applicable to multispectral images of natural environments.
References
Adams, J. B., Sabol, D. E., Kapos, V., Almeida, R. Filho, Roberts, D. A., Smith, M. O., Gillespie, A. R., Classification of multispectral images based on fractions of endmembers: Applications to land -use change in the Brazilian Amazon, submitted to Remote Sensing of Environment, in revision, 52:137-154, 1995.
Gillespie, A. R., Smith, M. O., Adams, J. B., Willis, S. C., Fischer, A. F. III, and Sabol, D. E., Interpretation of residual images: Spectral mixture analysis of AVIRIS images, Owens Valley, California, Proceedings Airborne Science Workshop: AVIRIS, Jet Propulsion Laboratory, Pasadena, CA, 4-5 June, 243-270, 1990.
Green, R. O., Conel, J. E., and Roberts, D. A., Es timation of aerosol optical depth and calculation of apparent reflectance from radiance measured by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) using MODTRAN2a, Proceedings Airborne Science Workshop: AVIRIS, Jet Propulsion Laboratory, Pasadena, CA, 25-29 October, 73-77, 1993.
Sabol, D. E., Roberts, D. A., Adams, J. B., and Smith, M. O., Mapping and monitoring changes in vegetation communities of Jasper Ridge, CA, using spectral fractions derived from AVIRIS images, Proceedings Airborne Science Workshop. AVIRIS, Jet Propulsion Laboratory, Pasadena, CA, 25-29 October, 157-160, 1993.
Sabol, D. E., Smith, M. O., Adams, J. B., Zukin, J. H., Tucker, C. J., Roberts, D. A., and Gillespie, A. R., AVIRIS spectral trajectories for forested areas of the Gifford Pinchot National Forest (abstract), submitted for Fifth Annual JPL Airborne Earth Science Workshop, January 23-26, 1995, 133-136, 1995.
Smith, M. O., Ustin, S. L. Adams, J. R., and Gillespie, A. R., Vegetation in deserts: II Environmental influences on regional abundance, Remote Sens. Environ. 31:27-52, 1990.