关键词: 多尺度空间注意力机制, 高斯核函数软标注, 华山松大小蠹（Dendroctonus armandi）, 受害木, 遥感识别

Abstract: To address the time-consuming and labor-intensive nature of traditional canopy boundary annotation, and the issue of decreased detection accuracy in existing deep learning models due to the loss of spatial details from downsampling in complex forest environments, a single-tree positioning method that integrated a multi-scale spatial attention mechanism convolutional network (MSSCN) and Gaussian kernel function soft labeling was proposed. Using high-resolution aerial remote sensing images from three altitude gradients (2 000 m, 2 200 m, and 2 400 m) in the Shennongjia Forestry District as the data source, only the canopy center points of Dendroctonus armandi were annotated. A two-dimensional Gaussian kernel function was employed to generate confidence maps for labeling and creating the training dataset, thereby transforming the regional segmentation task into a single-tree positioning problem. By adjusting the position of the multi-scale feature convolution module, MSSCN1, MSSCN2, and MSSCN3 models were constructed and compared with the U-Net, FCN, and DeepLabV3+ models. The results showed that the Gaussian kernel function soft labeling method reduced manual annotation costs while supporting the precise localization of infested trees. The MSSCN3 model achieved optimal performance after 100 training epochs, with average precision, recall, and F1-score values of 91.97%, 93.68%, and 0.93 in the test area, respectively, outperforming the other comparative models. The MSSCN3 model generally demonstrated superior detection performance in high-altitude areas of the Shennongjia Forestry District, and its detection accuracy was generally higher in high-outbreak-density areas than in low-outbreak-density areas. However, a slight decrease in model accuracy was observed in the high-outbreak-density area at 2 400 m altitude, indicating that topographic and ecological factors might have interactive effects on detection stability.The MSSCN3 model could accurately identify Pinus armandii infested trees in the Shennongjia Forestry District, providing an efficient and robust technical pathway for pest control.

Key words: multi-scale spatial attention mechanism, Gaussian kernel soft labeling, Dendroctonus armandi, infested trees, remote sensing identification