IV. LINEAR SPECTRAL UNMIXING |
Models designed to estimate class proportions (rather than a single class label) for individual pixels, while addressing the problem of mixed pixels and considering the spectral response from a mixture of classes, are referred to as mixture models [24], [25]. Such spectral mixture (SM) modelling was used on resampled 25 m Landsat-7 ETM+ data for a subpixel classification that achieved 87% accuracy for DN (Digital Number) values and 93% for radiance values [26]. The technique is useful for discerning information from data with low spatial resolution, and would thus be ideal for free MODIS data with large ground coverage [27]. The spectral radiance measured by the MODIS sensor consists of the radiances reflected by all materials present, thus the radiance can be summed in proportion to the sub-pixel area covered by each material, given that the endmembers are the reference spectra of each of the individual pure materials, and under the condition that these spectra are linearly independent. The sites of pure LC for each class (or component) of interest are identified, and their spectra are used to define endmember signatures as discussed in section III. The position of the spectral signature of an input pixel along this continuum indicates directly the percentage cover for each component [14]. Constrained Least-Squares method (CLSM) aids in computing n-1 variables with n simultaneous equations [19]. In the case of MODIS, seven bands were designed for LC mapping [28] and hence the maximum number of LC categories that can be obtained is only six.
Linear unmixing (and its variants - Multiple Endmember Spectral Mixture Analysis) has been used earlier for LC mapping using MODIS surface reflectance data of 250 m and 500 m spatial resolutions of Northern Africa [29] and the unmixed results were compared with high resolution classified maps that gave an overall classification accuracy of 54% with significant confusion between alluvial surfaces and regs (surface covering of coarse gravel/pebbles or boulders from which all sand and dust have been removed by wind and water; a stony desert), and between sandy and clayey surfaces and dunes. A second validation using 20 Landsat images in a stratified sampling scheme gave a classification accuracy of 70%, with confusion between dunes and sand sheets.
LC fractions derived from MODIS 1 km resolution data and MISR (Multi-angle Imaging SpectroRadiometer) using SM have been compared to results of a Bayesian-regularized artificial neural network (ANN), as well as with 30 m reflectance data, yielding a quantitative improvement over spectral unmixing of single-angle, multispectral data (Landsat-7/ETM+) [30].
