Deep Siamese Network with Handcrafted Feature Extraction for Hyperspectral Image Classification

The prominence of deep learning models for classification of hyperspectral images is directly proportional to their ability to exploit spatial context and spectral bands jointly. The effectiveness of these deep learning models, however, is heavily reliant on a good amount of labelled training samples. In contrast, one of the biggest challenges with hyperspectral images is limited labelled samples availability as getting the samples annotated is a time consuming and labor-intensive process. Traditional machine learning algorithms are available for classification with a higher training time and very deep pre-trained networks like GoogleNet and VGGNet did not work well for hyperspectral image classification. The idea of one shot classification has been quite motivating in recent years to deal with the problems of limited labelled samples, imbalanced distribution of samples leading to poor classification results and overfitting. To implement one shot classification model and overcome these challenges, the proposed work is based on Siamese network that can work with limited samples or imbalanced samples. The proposed Siamese network has a handcrafted feature generation network that extracts discriminative features from the image. Experimental findings on two benchmark hyperspectral datasets demonstrate that the proposed network is capable of improving the classification performance with an overall accuracy of 95.17 and 93.25 for Pavia U and Indian Pines dataset respectively with a small scale trained data.