Individual-tree identification of Korean pine based on UAV imagery and Mask R-CNN

In the Pinus koraiensis broadleaved mixed forest of Maoershan Experimental Forest Farm in Northeast China, we used a high-resolution orthophoto dataset acquired by a DJI Zenmuse P1 UAV to construct a dataset of individual tree crown recognition. Using Mask R-CNN, we extracted the individual-tree crown detection, boundary segmentation, and crown width extraction, which were compared with YOLOv9 under the same hardware conditions to clarify the differences in applicability between the two models in terms of recognition accuracy and inference efficiency. Then, we selected the P. koraiensis plantation in Hongqi Forest Farm and the P. koraiensis broadleaved mixed forest in Lushuihe Forestry Bureau as transfer-test sites, and conducted cross-forest-type transfer experiments based on the Mask R-CNN model trained on the Maoershan plot, to evaluate its usefulness under different stand conditions and its response to changes in training sample size. The results showed that, in the P. koraiensis broadleaved mixed forest of Maoershan Experimental Forest Farm, the mean average precision (intersection over union was 0.50), precision, recall, and F1 (the harmonic mean of precision and recall) of crown detection by Mask R-CNN were 0.83, 0.79, 0.82, and 0.80, respectively; and the corresponding values for crown boundary segmentation were 0.82, 0.78, 0.92, and 0.87, respectively. The crown width prediction achieved an R² of 0.89 and an RMSE of 0.42 m, and the overall accuracy was superior to that of YOLOv9. The inference speed of YOLOv9 was 63.7 FPS, approximately 4.3 times that of Mask R-CNN, making it more suitable for rapid large-area target localization. The transfer experiments showed that the initial detection and segmentation accuracies of Mask R-CNN in the P. koraiensis plantation of Hongqi Forest Farm were 0.82 and 0.80, respectively, indicating relatively stable performance. In the P. koraiensis broadleaved mixed forest of Lushuihe Forestry Bureau, the initial detection and segmentation accuracies were 0.44 and 0.42, respectively. Although the initial accuracy was relatively low, the improvement became more pronounced as the number of training samples increased. In summary, Mask R-CNN showed better applicability in the recognition of individual P. koraiensis tree, crown boundary segmentation, and crown width estimation, whereas YOLOv9 had advantage in rapid large-area inventory scenarios. This study provided technical support for P. koraiensis resource inventory, structural parameter extraction, and dynamic monitoring in Northeast China.