Exploiting spatial correlation of spectral signature for training data selection in hyperspectral image classification

Supervised classification is commonly used to produce a thematic map from hyperspectral data. A classifier is learned from training pixels and used to assign a known class (theme) to each pixel (imagery data example). However, supervised classification requires a sufficient number of representative training samples to be accurate. These samples are usually selected by expert visual inspection or field survey. Consequently, collecting representative samples is a very challenging task due to the high cost of true sample selecting and labeling. This paper introduces an unsupervised learning schema, where the most suitable pixels to train the classifier are selected via image segmentation. This reduces the expert effort required for choosing training samples. In our proposal, clustering is performed by accounting for the property of spatial correlation of pixel-level spectral information, so that thematic objects can be retrieved via unsupervised learning and representative training data can be sampled throughout clusters. Experimental results highlight that the pixel classification accuracy outperforms the results of a random selection scheme.