Reduction of Feature Contamination for Hyper Spectral Image Classification

Abstract

Motivated by the power of the contrastive learning process, in this paper we present a novel supervised contrastive learning network add-on which reduces the misclassifications of the state-of-the-art Hyper Spectral Image (HSI) classification models. We observe that a significant number of misclassification of these HSI classification models occur at the class borders where there exist multiple different classes in the neighbourhood. We believe this is due to the contamination of feature space in the deeper layers of the CNN network. To mitigate this deficiency we propose a novel supervisory signal design that ‘pulls' the features derived from the same class as of class of the centre pixel together, while ‘pushing’ the features of other classes far apart. This yields a novel trainable neural network module for Reducing Feature Contamination (RFC). The proposed module architecture is model agnostic and can be coupled with different CNN based architectures where it is required to alleviate the contamination of spectral signatures from neighbouring pixels of other classes. Through extensive evaluations using the state-of-the-art SSRN, HybridSN, A2S2K-ResNet and WHU-Hi, Indian Pines and the PaviaU datasets we have demonstrated the utility of the proposed RFC module.