Unmixing multiple land-cover type reflectances from coarse spatial resolution satellite data

Abstract

We present a technique for extracting subpixel cover type reflectances from the mixed pixels of a coarse spatial resolution image. A Gaussian filter is used to spatially degrade a set of fine spatial resolution images [based on, e.g., Landsat Thematic Mapper (TM)], each of which represent the spatial structure and extent of a land-cover type. From the degraded image data, a set of weights representing the proportions of cover types within the mixed pixels of a coarse spatial/fine temporal resolution image [e.g., NOAA Advanced Very High Resolution Radiometer (AVHRR)] is produced. These weights and the AVHRR spectral band reflectances are used in multiple linear regression analyses to extract mean cover type reflectances. A simulation study was conducted to define the operational limitations of this technique. Cover type reflectances are accurately estimated if the cover types are spatially well represented within the area of interest, are spectrally distinct, and have small within-class spectral variability. The accuracy of retrieved reflectances is most sensitive to errors in the coarse spatial resolution data and least sensitive to errors in the weights. By applying this technique to multiple images, temporal profiles of subpixel cover type reflectance may be obtained. This particular application is demonstrated by using TM and AVHRR data from the First International Satellite Land Surface Climatology Project (ISLSCP) Field Experiment (FIFE) to create normalized difference vegetation index (NDVI) temporal profiles of the dominant land-cover types.