Spectral unmixing and sub-pixel classification: Analysis of learning strategies

Sub-pixel details in the hyperspectral images are generally ignored by the conventional classifiers. However, some recent approaches use this information to generate fine resolution land cover maps from images having coarse spatial resolution. Two main aspects in this regard are: 1) estimation of fractional abundances of the reference signatures at each pixel (spectral un-mixing); and 2) prediction of class distributions at sub-pixel scale (sub-pixel classification). This study proposes some spectral unmixing as well as sub-pixel mapping techniques that take in to account certain constraints which are usually ignored by the conventional approaches. In the context of spectral unmixing methods, our main contribution is the analysis of auto-encoders when compared with ELM, STM and SVM. In case of sub-pixel mapping methods, our study may be summarized as the modelling deep auto-encoders for predicting the spatial distributions at target scale. Also, we have compared the effectiveness of Auto-Encoders and their convolutional counterparts in learning the coarse image features. Among the proposed unmixing approaches, autoencoder approach has given better results when compared to that of SVM and STM. The deep learning based sub-pixel mapping approaches have also produced good results, even for complex scenes. The sensitivities of all these techniques towards various tunable parameters are also analyzed.