Multi-scale vision transformer model with UAV-based hyperspectral imagery for early identification of pine wilt disease infected trees

Pine Wilt Disease (PWD), caused by the pine wood nematode ( Bursaphelenchus xylophilus ), is a devastating forest disease characterized by rapid spread and high mortality. During its pathogenesis, the green attack and early stage are critical for disease control. However, RGB UAV images lack obvious visual features at these stages, which limits the accuracy of existing identification methods. This study aims to develop a Multi-Scale Vision Transformer Model specifically designed for UAV hyperspectral data. The primary objectives are to improve the identification capability for the green attack and early stages of PWD. To accurately capture subtle changes in the spectral and spatial features of pine trees during the green attack stage, an efficient spectral-spatial feature fusion module was designed. Meanwhile, an improved Gaussian self-attention mechanism was introduced to enhance the modeling of spectral differences between central pixels and their neighboring pixels. Furthermore, by integrating a lightweight Transformer architecture with a multi-scale feature pyramid, the proposed method not only fuses features ranging from the pixel level to the canopy level but also reduces the number of parameters. Experiment results show that the model achieves an identification accuracy of 87.40% and 88.21% for the green attack stage and the early stage of PWD, respectively. The model achieves an OA of 87.90% and a Kappa coefficient of 83.25%, both of which outperform those of traditional machine learning methods and other deep learning models. In terms of efficiency, the model has only 2.45 M parameters, a FLOPs of 112.42 M, a training time of 114 min, and a testing time of 25 min, demonstrating high efficiency. The model improves the recognition accuracy for the green attack stage and early stage through targeted design, providing important technical support for the large-scale monitoring and scientific prevention and control of PWD-infected trees.