Sensitive Band Selection Method for Hyperspectral Images of Pine Wilt Disease Infected Trees Based on Adaptive Subspace Partitioning

Pine wilt disease (PWD), referred known as “pine cancer,” is incurable and highly contagious, with a high mortality rate, posing a serious threat to global pine resources. Early detection, localization, and clearance of infected trees remain the only effective control strategies. However, the subtle physiological alterations occurring in early-stage infected trees are scarcely discernible in visible-light imagery, rendering detection through conventional computer vision techniques extremely challenging. This study demonstrates the feasibility of hyperspectral remote sensing for the early detection of PWD-infected trees. Targeting the weak pathological spectral responses characteristic of early infection, we develop a novel adaptive subspace partitioning (ASP) mechanism that adaptively identifies pathology-sensitive spectral bands while reducing hyperspectral data redundancy. This approach enables the refined extraction of subtle but diagnostically significant spectral cues, providing a new methodological framework for early detection of PWD-infected trees. Furthermore, to enhance spectral feature learning, a spectral reconstruction network integrating a multi-attention mechanism with a CNN–ViT architecture is introduced, enabling joint mining of spatial–spectral contextual information and precise modeling of infected-tree spectral properties. By combining the spectral reconstruction network with the ASP mechanism, the PwdBSNet model for hyperspectral band selection of PWD-infected trees is constructed. Experimental results demonstrate that, PwdBSNet consistently outperforms mainstream band selection methods including GPCBS and FNGBS. Compared with baseline models, PwdBSNet achieves improvements of 8.35% in OA, 9.91% in AA, and 0.11 in Kappa, offering a novel and effective approach for hyperspectral band selection in PWD detection.