Dirichlet process based context learning for mine detection in hyperspectral imagery

Abstract

Hyperspectral imagery (HSI) has been shown to be a powerful remote sensing phenomenology that is appropriate for a variety of classification and detection tasks. Standard detection and classification algorithms applied to hyperspectral data are hindered by environmental factors that alter the statistics of the data such as sun intensity, atmospheric conditions or soil properties. Detection and classification algorithms operating on HSI must account for the changing context underlying each observation for robust performance. This work focuses on algorithms that incorporate knowledge of underlying context for the discrimination of landmine responses from other surface or sub-surface anomalies using airborne HSI. This work compares both generative context models, that model context at a given location using features of the surrounding data, and discriminative context models that determine the context at a given location to maximize performance. Both approaches utilize a Dirichlet process prior to infer the number of contexts within the data without the need to explicitly label the context of each image or location within the image. Results indicate that Dirichlet process based generative context clustering determines contexts that are congruent with physical characteristics such as time of day, but does not necessarily lead to performance improvements. Dirichlet process based discriminative clustering, however, yields performance greater than a labeled generative approach.