Hyperspectral Image Classification Using Denoised Stacked Auto Encoder-Based Restricted Boltzmann Machine Classifier

This paper proposes a novel solution using an improved Stacked Auto Encoder (SAE) to deal with the problem of parametric instability associated with the classification of hyperspectral images from an extensive training set. The improved SAE reduces classification errors and discrepancies present within the individual classes. The data augmentation process resolves such constraints, where several images are produced during training by adding noises with various noise levels over an input HSI image. Further, this helps in increasing the difference between multiple classes of a training set. The improved SAE classifies HSI images using the principle of Denoising via Restricted Boltzmann Machine (RBM). This model ambiguously operates on selected bands through various band selection models. Such pre-processing, i.e., band selection, enables the classifier to eliminate noise from these bands to produce higher accuracy results. The simulation is conducted in PyTorch to validate the proposed deep DSAE-RBM under different noisy environments with various noise levels. The simulation results show that the proposed deep DSAE-RBM achieves a maximal classification rate of 92.62% without noise and 77.47% in the presence of noise.