Classification of Hyperspectral Remote Sensing Images Using High-level Features Based on Empirical Modes

Abstract

The role of empirical mode decomposition (EMD) in the synthesis of high-level features for the classification of hyperspectral remote sensing images is studied. The studies were performed on the material of the well-known HSI "Moffett Field". A 1D-EMD algorithm adapted to the needs of HSI analysis was used. It has been established that: 1) class reference in the form of only a reference HSI-signature of a class sample cannot be a sufficient feature for classification on the full "Moffett Field" HSI; 2) the extention of an HSI-object class standard, consisting of a reference signature (spectral characteristic) of a class sample, even by one of the first empirical modes, either dramatically increases the contrast between the standards, or reveals the indistinguishability of the standards in the global coordinate system (belonging to the same class); 3) empirical modes are able to provide the necessary refinement of the class standard for a variety of HSI Moffett Field objects; 4) formation rules for a high-level spectral feature from empirical modes are proposed.