Integrating principal component analysis in spatial-spectral fusion models for hyperspectral image segmentation

Abstract

Hyperspectral imaging (HSI) from unmanned aerial vehicles (UAVs) provides rich spatial-spectral data, but its high dimensionality presents significant computational challenges for semantic segmentation. While state-of-the-art models like the transformer-based HSI-TransUnet are often employed, they introduce massive computational overhead. This study adapts a lightweight, dual-tunnel deep convolutional neural network (DCNN) framework for land-use segmentation on hyperspectral images by integrating PCA-based spatial reduction in the spatial branch, and benchmarks it on the UAV-HSI-Crop dataset against HSI- TransUnet. For further analysis, an ablation study compares principal component analysis (PCA) and local similarity projection (LSP) as spatial feature ex- tractors. The results demonstrate a significant performance and efficiency advantage. Our proposed PCA-based model (271.1K parameters) obtained a Kappa (κ) of 0.8582, overall accuracy (OA) of 0.8800, and average accuracy (AA) of 0.4918, outperforming the LSP-based model by 0.65% in κ, 0.51% in OA, and 2.16% in AA and the HSI-TransUnet baseline by 2.35% in κ, 1.95% in OA, and 8.10% in AA. On our experimental setup, this result was achieved with a 152.7-fold reduction in model size, a 14.2-fold decrease in training time, and a 4.6-fold speedup in inference relative to the reported HSI-TransUnet baseline. These findings show that the PCA-based dual-tunnel DCNN provides a favor- able trade-off between class-balanced accuracy and computational efficiency for this HSI segmentation task.