Uncertainty Assessment of Spectral Mixture Analysis in Remote Sensing Imagery

Spectral mixture analysis (SMA), a scheme of sub-pixel-based classifications, is one of the widely used models to map fractional land use and land cover information in remote sensing imagery. It assumes that: 1) a mixed pixel is composed by several pure land cover classes (endmembers) linearly or nonlinearly, and 2) the spectral signature of each endmember is a constant within the entire spatial extent of analysis. SMA has been commonly applied to impervious surface area extraction, vegetation fraction estimation, and land use and land cover change (LULC) mapping. Limitations of SMA, however, still exist. First, the existence of between- and within-class variability prevents the selection of accurate endmembers, which results in poor accuracy of fractional land cover estimates. Weighted spectral mixture analysis (WSMA) and transformed spectral mixture analysis (TSMA) are alternate means to address the within- and between- class variability. These methods, however, have not been analyzed systematically and comprehensively. The effectiveness of each WSMA and TSMA scheme is still unknown, in particular within different urban areas. Second, multiple endmember SMA (MESMA) is a better alternative to address spectral mixture model uncertainties. It, nonetheless, is time consuming and inefficient. Further, incorrect endmember selections may still limit model performance as the best-fit endmember model might not be the optimal model due to the existence of spectral variability. Therefore, this study aims 1) to explore endmember uncertainties by examining WSMA and TSMA modeling comprehensively, and 2) to develop an improved MESMA model in order to address the uncertainties of spectral mixture models. Results of the WSMA examination illustrated that some weighting schemes did reduce endmember uncertainties since they could improve the fractional estimates significantly. The results also indicated that spectral class variance played a key role in addressing the endmember uncertainties, as the better performing weighting schemes were constructed with spectral class variance. In addition, the results of TSMA examination demonstrated that some TSMAs, such as normalized spectral mixture analysis (NSMA), could effectively solve the endmember uncertainties because of their stable performance in different study areas. Results of Class-based MEMSA (C-MESMA) indicated that it could address spectral mixture model uncertainties by reducing a lot of the calculation burden and effectively improving accuracy. Assessment demonstrated that C-MEMSA significantly improving accuracy. Major contributions of this study can be summarized as follow. First, the effectiveness of addressing endmember uncertainties have been fully discussed by examining: 1) the effectiveness of ten weighted spectral mixture models in urban environments; and 2) the effectiveness of 26 transformed spectral mixture models in three locations. Constructive guidance regarding handling endmember uncertainties using WSMA and TSMA have been provided. Second, the uncertainties of spectral mixture model were reduced by developing an improved MESMA model, named C-MESMA. C-MESMA could restrict the distribution of endmembers and reduce the calculation burden of traditional MESMA, increasing SMA accuracy significantly