Conclusion

In general, most cases of sub-pixel analysis assume that classes may be described by unique endmembers. However, due to class spectral variability, no single spectrum can describe a class adequately. As a result, distribution of possible class compositions for each pixel exists, where accuracy of sub-pixel class abundances are functions of intra-class spectral variations. In this communication, we proposed a new method – variable endmember constrained least square (VECLS) which has the ability to incorporate the intra-class variability of the spectral signatures in the unmixing process. The covariance matrix for each endmember obtained from the original bands was accounted in classification for the variability in each class. The results of the algorithm implementation on computer simulated data with small, medium and large variability resulted in an average correlation increment of 0.25. The application on RS data such as IKONOS, Landsat ETM+ and MODIS also showed promising results with increased average correlation for different number of classes compared to absence of endmember variability, demonstrating the effectiveness of this approach.

Acknowledgement

We are grateful to Indian Institute of Science for financial and infrastructure support. Global Land Cover Facility (GLCF) is appreciated for providing the Landsat and MODIS data. We thank Geo Eye Foundation, USA for providing IKONOS imagery for Greater Bangalore City.