Semisupervised Hyperspectral Image Classification Using Small Sample Sizes

摘要

Hyperspectral image classification is a challenging task when only a small number of labeled samples are available due to the difficult, expensive, and time-consuming ground campaigns required to collect the ground-truth information. It is also known that the classification performance is highly dependent on the size of the labeled data. In this letter, a semisupervised learning-based hyperspectral image classification framework is proposed as a solution to these problems. One of the contributions of this letter is the selection of the initial labeled training samples with a subtractive clustering-based approach, which provides the most informative samples for graph-based self-training. Another contribution is the decision-level combination of results obtained by support vector machines and kernel sparse representation classifiers. Additionally, a combination of the spatial and spectral information by creating a window structure is also proposed via integrating contextual information from the neighboring pixels. The explanatory experiments confirm that the proposed framework offers better and more promising results, even using a small number of initial labeled samples.