Sampling training data for accurate hyperspectral image classification via tree-based spatial clustering

The classification of hyperspectral images is a challenging task due to the high dimensionality of the task (i.e. large amount of pixels described over a high number of spectral channels) coupled with the small number of labeled examples typically available for learning. In the last decades, Support Vector Machines (SVMs) have gained in popularity in the field of the hyperspectral image classification as they address large attribute spaces and produce solutions from sparsely labeled data. However, they require “representative” training samples of the unknown class distribution to be accurate. In general, these samples are manually selected by expert visual inspection or field survey. This paper describes a learning schema, where the most suitable pixels to train the classifier are automatically selected via a spectral-spatial clustering phase. This reduces the expert effort required for sampling training pixels. Experimental results highlight that the proposed solution allows us to achieve a classification accuracy that outperforms the accuracy of both random and baseline sampling schemes.