A comparison of mixture modeling algorithms and their applicability to the MODIS data

The performance of three different mixture models is tested on two sets of synthetic data which were generated by spatially degrading a Landsat-TM image. The first data set was generated by using a simple weighted spatial averaging filter, whereas the second data set was generated by applying the point spread function (PSF) of the Moderate Resolution Imaging Spectroradiometer (MODIS). Two of the models (MIXWLS and UNMIX) solve the mixture decomposition problem by statistical least square methods, and require a priori information about the spectral signatures of the individual end-members. The third model (COCKTAIL) identifies the end-member spectral values from the image itself through factor analysis. All the three methods showed larger errors in their estimates of the proportion of end-member in each pixel in the MODIS data compared to those from the spatially averaged data due to PSF effects. However, the magnitude of errors was surprisingly similar from all the three models when the simulated MODIS data was used.