A random measure approach for context estimation in hyperspectral imagery

Abstract

In remotely sensed hyperspectral imagery (HSI), images are collected in the presence of various contextual factors which change the distribution of the observed data. Examples of these factors are suns intensity, atmospheric constituents, soil moisture, local shading, etc. In this paper, a context based classification algorithm is developed which implicitly identifies context without explicitly needing environmental data (as in may be unknown or locally variable). Spectra sets are clustered into groups of similar contexts using a random measure model. Then appropriate classifiers are constructed for each context. The resulting context-based classification algorithm constructed within the random set framework then aggregates the classifiers results in an ensemble-like fashion. Results indicate that the proposed approach performs well in the presence of contextual factors.