A Modified Fully Convolutional Segmentation Network for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is pivotal for precise land-cover analysis and environmental monitoring. Methods based on Convolutional Neural Networks (CNNs) in deep learning have garnered interest in HSI classification (HSIC). However, they often struggle with poor generalization due to complex spatial land-cover patterns in HSI data. The Fully Convolutional Segmentation Network (FCSN) enhanced the HSI classification’s generalization capabilities. Yet, FCSN faced accuracy loss on original datasets with limited labeling, particularly in low-labeled environments. The FCSN overfits to completely labeled training samples. To address this issue, we proposed a modified Fully Convolutional Segmentation Network (mFCSN), preserving generalization capabilities while greatly enhancing accuracy even with minimal labeled HSI data. The architecture involves two parallel CNN paths: the primary, with multiple residual connections for network convergence, and the secondary, with added filters for extracting spatial-spectral features. This dual-parallel-path strategy enables simultaneous exploration of diverse hyperspectral data aspects. The proposed mFCSN achieved superior performance, for instance, attaining an overall accuracy of 77.19% on the Indian Pine dataset and remaining unaffected by changes in spatial patterns.