Applying spectral mixture analysis for large-scale sub-pixel impervious cover estimation based on neighbourhood-specific endmember signature generation

Spectral mixture analysis (SMA) has been extensively adopted in estimating sub-pixel impervious surface fractions. As a key step of SMA, endmember extraction has a big impact on the reliability of unmixing result. Due to the difficulty in extracting spectrally pure pixels using traditional methods, SMA is seldom applied to coarse-resolution imagery. A promising strategy to overcome this challenge is to synthesize endmember signatures via generalized least squares solution (LSS) technique with known fractions of samples. However, this method yields constant endmember spectra across the entire image extent, indicating a potential over simplification of spatial heterogeneity. As such, in this study we developed a neighbourhood-specific endmember signature generation method to derive spatially variable endmember signatures using geographically weighted regression technique. According to our investigation results, the developed method performed well in mapping fractional imperviousness with a single date Moderate Resolution Imaging Spectroradiometer imagery and exhibited relatively high estimation accuracy (root mean square error of 10.98%, mean absolute error of 8.45% and bias of 0.25%) compared with the generalized LSS method.