Controlled spectral unmixing using extended Support Vector Machines

Abstract

This paper presents an improved spectral unmixing framework for remote sensing data interpretation. Instead of unmixing every pixel in an image into a fixed set of endmembers, approaches of adaptive subsets of endmember selection for individual pixels are presented which can improve the performance of spectral unmixing. An integrated hard and soft classification map is then generated by applying the mixture analysis based on extended Support Vector Machines. The proposed treatment is effective and easy to implement. Unmixing is more reliable with the controlled mixture model. It can cope with the endmembers' spectral variation as a result of system noise encountered during data collection from the space. Experiments were conducted with Landsat ETM data and satisfactory results were achieved.