Hyperspectral Image Classification Based on the Lightweight G-GhostNets Network

Hyperspectral image classification (HSIC) has been widely applied in remote sensing image analysis. However, deploying deep learning models on embedded devices for HSIC tasks faces challenges due to excessive parameters and high computational complexity. To address these issues, this paper proposes the lightweight G-GhostNets network optimized for GPU devices. G-GhostNets effectively integrates both spectral and spatial features of hyperspectral images. By introducing the G-Ghost Stage, which aggregates intermediate features, the model achieves high-precision classification while maintaining computational efficiency on GPUs. To validate its performance, this study used ResNet50 and C-GhostNets as comparative networks and conducted classification experiments on preprocessed Pavia University, Indian Pines, and Salinas datasets, respectively. The experimental results demonstrate that G-GhostNets exhibits outstanding performance in HSIC tasks, delivering high accuracy and efficient GPU-based inference with significantly reduced computational time per class.