Self-organizing maps for fusion of thermal hyperspectral- with high-resolution VIS-data

A new hyper-spectral data set is at hand giving unique possibilities for investigating also multi-scale evidence fusion. In this contribution self-organizing maps are used for semi-supervised learning and visualization of the partially labeled data. The maps reveal that the seven classes given can be better distinguished using certain color and rotationally invariant texture features on the high-resolution visual data than on the thermal spectral data. These spectra are very similar to each other. Still the self-organization can also elaborate subtle differences that exhibit some discriminative possibilities. Best class separation results from fusion of both sources. But of course more computational effort is needed, and convergence is slower due to the higher dimensionality of the fused feature space.