Objective The UAV image automatically extracts canopy for precise information.

Method In this study, single tree crown extraction method from UAV images based on Mask R-CNN and transfer learning was proposed. Firstly, the optimized Mask R-CNN instance segmentation model based on faster R-CNN was selected, and the optimal feature extraction network was chosen between ResNet50 and ResNet101. Secondly, transfer learning was introduced to train with Mask R-CNN together, combined with the guiding role of transfer learning to reduce training time and improve training accuracy.

Result The results showed that the overall accuracy of the Mask R-CNN model was 93.59%, the user accuracy, F1 score and mean average precision were 65.46%, 76.05% and 0.31, respectively. After adding transfer learning, the user accuracy of Mask R-CNN model was increased by 29.53%, the F1 score was increased by 19.63%, and the average accuracy was increased by 0.21. In the Mask R-CNN model with ResNet50 and ResNet101 as feature extraction networks, the average values of overall accuracy, user accuracy, F1 score and mean average precision of the ResNet50 + Mask R-CNN model were 96.94%, 95.57%, 96.17% and 0.54, respectively. The average values of overall accuracy, user accuracy, F1 score and mean average precision of the ResNet101 + Mask R-CNN model were 96.20%, 94.41%, 95.19% and 0.49, respectively. The R² of ResNet50 + Mask R-CNN model loaded with transfer learning in predicting east-west crown width, north-south crown width, crown area and quadrat canopy density were 0.87, 0.84, 0.93 and 0.83, respectively.

Conclusion The method based on Mask R-CNN and transfer learning proposed in this study obtains more accurate results of tree crown parameters, which provides a fast and efficient solution for tree resource assessment in UAV images.