Multiple instance and context dependent learning in hyperspectral data

Hyperspectral imaging (HSI) is a powerful tool for various remote sensing tasks including agricultural modeling and landmine/ unexploded ordnance clearance. Although the application of standard supervised learning techniques to HSI data has previously been explored, several aspects of hyperspectral data collection and ground truth labeling make some of the assumptions underlying standard machine learning techniques invalid. For example, HSI is highly dependent upon local environmental conditions, and pixel-by-pixel labels for HSI data are often not available. As a result, data from hyperspectral sensing under various scenarios is not typically i.i.d., and correct data labels must be inferred from training data while learning decision boundaries. In this work we explore two possible solutions to these problems: context-dependent learning for overcoming variations between collections, and multiple instance learning for simultaneously inferring local target labels and global target decision boundaries. Results are compared to standard logistic discriminant classification approaches.