A comparison of principal components and endmember-based contextual learning for hyperspectral anomaly classification

Context-dependent learning algorithms have shown improved performance for anomaly classification in hyperspectral imagery (HSI) collected over varying environmental conditions. Past techniques have relied on statistically-motivated decomposition, such as principal components analysis (PCA), to extract contextual information from the background data. Alternatively, physics-based endmember approaches could also be used to extract contextual features. In this work, context-dependent classifiers using both types of contextual features were applied to a landmine detection problem in HSI. Context-dependent learning showed improvements in classification performance over conventional learning, and the endmember-based and PCA-based context modeling techniques yielded similar overall model behavior which is investigated.