Considerations on unsupervised spectral data unmixing and complexity pursuit

Hyperspectral sensors carry the distinctive advantage of recording hundreds of contiguous spectral images for the same scene providing an extraordinary amount of information that leads to precise differentiation of materials present in the scene even when such materials contribute only to few pixels. With the advent of more and more powerful sensing platforms, coupled with reduction in manufacturing costs and diversification of technologies, hyperspectral imaging has become a powerful approach in remote sensing with applications spanning all traditional fields (such as agriculture, mining, military, resource management, etc.) as well as new ones (manufacturing quality control, pollution detection, health and life sciences, food safety etc.) In this paper we tackle the complexity based unmixing and develop new techniques that generalize the concept of spatial complexity to larger neighborhoods. Furthermore we assess the value of spatial complexity pursuit for small targets. Finally, we align the complexity based model with the linear mixing model by including additional conditions such as positivity and additivity.