SMNDO-PWD: A Density-Adaptive Otsu Framework for Cross-Regional Detection of Pine Wilt Disease Using RGB Imagery and Spectral Reconstruction

Pine wilt disease (PWD) has caused extensive damage to pine forests in China. However, large-scale monitoring using high-resolution imagery is constrained by the high cost of hyperspectral data and the annotation requirements of task-specific deep learning models. To address these challenges, this study proposes SMNDO-PWD, a remote sensing framework for detecting PWD-affected trees. First, a pretrained SegFormer (MiT-B5) model removes non-vegetation features such as buildings and bare soil. MST++ then reconstructs statistically plausible hyperspectral bands in the 400–700 nm range from the RGB imagery, and the normalized difference index (NDI) is computed from the reconstructed 670 nm and 550 nm bands. Subsequently, support vector machine (SVM)-based density classification is integrated with density-adaptive dynamic Otsu thresholding to extract PWD-affected trees, requiring only lightweight classifier training and threshold optimization. Results from three reference plots in Shaoguan show that SMNDO-PWD achieves average pixel-level intersection over union (IoU) and F1 scores 51.28% and 32.14% higher, respectively, than YOLOv8s-seg. Further evaluation across eight cities in six provinces yields average IoU and F1 scores of 0.52 and 0.69, respectively, using manual annotations as reference. Ablation experiments confirm the contributions and complementary effects of individual modules. Furthermore, an application covering approximately 900 ha in Zhaoqing demonstrates the potential of SMNDO-PWD for large-scale monitoring. The framework also enables the quantification of the affected area ratio (AAR), affected area, and the relative NDI distribution in the affected areas, providing spatial information to support PWD assessment and management.