News

We consider the problem of designing features for the identification of reflectance spectra that capture the semantic information used by experts in ad-hoc labeling rules, such as the shape and position of absorption bands in spectra. We propose the use of statistical models on wavelet coefficients that allow us to quantify the presence of spectrum discontinuities at multiple scales.

We consider the problem of designing features for the identification of reflectance spectra that capture the semantic information used by experts in ad-hoc labeling rules, such as the shape and position of absorption bands in spectra. We propose the use of statistical models on wavelet coefficients that allow us to quantify the presence of spectrum discontinuities at multiple scales. Such models are now feasible to train due to the availability of large-scale datasets of hyperspectral images. Using a non-homogeneous hidden Markov model, we can succinctly express the semantic information in the spectrum in terms of binary state labels representing its energy content in a multiscale time-frequency analysis. Experimental results show that the features succeed in outperforming existing approaches in encoding and quantifying the semantic features of the spectra that are relevant in classification tasks. This is joint work with Marco Duarte and will be presented at the IEEE Workshop on Hyperspectral Imaging and Signal Processing - Evolution in Remote Sensing (WHISPERS), June 25-28, 2013 in Gainesville, FL