An evaluation of spectral mixture modelling applied to a semi-arid environment

The number of ground cover classes for which proportions estimates can be derived by application of spectral mixture modelling is constrained by the dimensionality of the data. This may limit the usefulness of the approach, as often there are a larger number of discriminated classes than dimensions of data. Typically, relatively few, broadly defined end-member classes have been used but this is likely to degrade the precision of the proportions estimates at the pixel level as the end-member signatures attempt to represent a spectrally diverse range of vegetation covers within each broad class. To overcome this problem, multiple end-member mixture models have been proposed that iterate through a larger number of more narrowly defi'ned end-members. In this study, broad and multiple end-member mixture models were applied to satellite data at 1 km(2) spatial resolution of a semi-arid environment to evaluate their respective precision of proportions estimates against ground survey data for 25 1-km(2) segments. This study demonstrates improved precision of proportions estimates by application of a multiple end-member mixture model. However, the study also shows that the achievable level of precision of proportions estimates using mixture modelling for a semi-arid environment and satellite data can range within limits of 0.22 to 0.84. These results are unlikely to be acceptable for environmental monitoring purposes and the technique cannot be deemed a suitable approach for the study of semi-arid vegetation cover types using mixture model analysis of remote sensing data with intermediate spatial resolution.