A MASSIVE SELF-ORGANIZING MAP FOR HYPERSPECTRAL IMAGE CLASSIFICATION

Hyperspectral image classification is a challenging task due to its high dimensionality. Class imbalance and lack of enough sample images make the process of classifying hyperspectral remote sensing images (RSI) more difficult. One well-known unsupervised learning algorithm called self-organizing maps (SOM) which utilizes competitive learning for training neurons to represent particular subsets of a dataset well. The SOM has been used for classification with RSI but does not generate any outstanding results. In addition, by using the typically small map sizes for SOM, the minority classes are not recognized adequately. To improve SOM, the concept of Emergent SOM (ESOM) was introduced where the use of much larger map sizes of at least four thousand neurons was recommended. To handle the class imbalance in RSI with unsupervised classification, we propose the use of much larger and more massive SOM than suggested. With a more significant number of representable cluster centers, the minority classes can be detected as the projection onto a map size of that magnitude magnifies the input subset resulting in a higher granularity with a deeper level of detail. Our experimental results show that the massive SOM (MSOM) can classify entire hyperspectral images with impressive classification accuracy along with identifying minority classes accurately. Simultaneously, the performance of MSOM rivals current contemporary supervised learning methods.