Using AI and deep learning techniques for tree species recognition in UAV forest aerial images is a valuable research domain. Since different network models may produce different prediction results when performing tree species recognition, using ensemble algorithm to combine the prediction results of multiple models can overcome single model's limitation and effectively improve the overall accuracy. However, when there are many candidate models, using all of them to do prediction respectively and combining the results will take too many time. More efficient way is to select top n models among them and only use the chosen models to do prediction and combine their results. In this study, we collected 3 km2 of high-definition UAV forest aerial images and labeled 9500+ samples, including bamboo, pine and chinese fir 3 species trees, as the dataset. To select the top n “best” models, we proposed a novel approach considering both the accuracy and the diversity of prediction results among different models. The top n models with the highest comprehensive scores were selected from all candidate models and the results were combined using the Weighted Boxes Fusion algorithm. The experimental results showed that compared to the strategy of combining all candidate models, the proposed strategy reduced the average recognition time from 1.72 seconds to 0.78 seconds (when n=3) and 0.503 seconds (when n=2) respectively. Compared with the combination strategy that only considers each model's prediction accuracy, our strategy had a significant improvement in both overall accuracy and mAP.
1 January 2026 in IEEE Geoscience and Remote Sensing Letters