A Pioneering Approach of Hyperspectral Image Classification Employing the Cooperative Efforts of 3D, 2D and Depthwise Separable-1D Convolutions

Abstract

Due to its promising uses in diverse sectors including agriculture, food technology, land cover mapping, etc., hyperspectral image (HI) analysis plays an important role. HI analysis using deep learning (DL) has made enormous strides in recent years. The literature makes it clear that employing simply 2-D-CNN or 3-D-CNN included several limitations, such as inadequate channel association details or mandating very complex models, respectively. The most important justification is that HIs are volumetric data with several spectral bands. This paper proposes a synergetic effort by 3D, 2D, and Depthwise separable-1D Convolutional layers in the proposed Lightweight Deep Learning model for enhancing HI classification accuracy. The spectral dimension of the input HI is reduced by utilizing traditional Principal Component Analysis (PCA). The 3D CNN expedites the synthesis of simultaneous spatial-spectral characteristics out of a pile of spectral channels. The 2D CNN also picks up the higher-level spatial interpretation in addition to the 3D CNN. Finally, the Depthwise separable-1D Convolutional layer that works on depth and space is incorporated into the model. The introduction of the Depthwise separable-1D Convolutional layer, the quantity of the channels, and the kernel sizes resulted in a sharp reduction in computation. Strenuous classification trials are conducted over the Indian Pines, University of Pavia, and Salinas Scene HI data to assess the viability of this proposed approach. The findings are juxtaposed with cutting-edge hand-crafted and sophisticated deep learning-based methods. The suggested DL architecture for HI classification performs quite effectively with fewer training samples and less trainable parameters.