MsGsFNet: A Multi-Scale Spatial and Global Spectral Information Fusion Network for UAV-H2 imagery classification

Unmanned aerial vehicles (UAVs) equipped with hyperspectral imaging systems can capture high spatial and spectral resolution (H²) imagery, providing detailed information about ground objects. However, high spatial and spectral resolutions increase spatial heterogeneity and spectral variability, which pose challenges for accurate land cover classification. To address these problems, we introduce a global spectral feature extraction module that integrates non-negative matrix factorization (NMF) and an intraclass channel weight cosine distance constraint to reduce spatial–spectral redundancy and enhance class distinction. Additionally, we propose a multi-scale spatial feature adaptive extraction module using large kernel convolutions to reduce the impact of spatial heterogeneity on classification accuracy. Re-parameterization techniques are also employed to improve inference speed without increasing computational costs. Based on these designs, we developed an efficient spatial-spectral information fusion classification network (MsGsFNet). To objectively evaluate its performance, we designed a novel spatially disjoint sampling method to avoid data leakage risks from random sampling. Experimental results on four public datasets demonstrate that MsGsFNet consistently achieves state-of-the-art performance, with 92.10% OA and 70.53% MIoU on WHU-HI-HongHu, 99.15% OA and 96.26% MIoU on WHU-HI-LongKou, 91.37% OA and 77.75% MIoU on WHU-HI-HanChuan, and 93.18% OA and 81.31% MIoU on QUH-QingYun. The re-parameterization technique also boosts FPS by 588% on WHU-HI-HongHu. Our code will be available at https://github.com/CHD-IPAC/MsGsFNet.