Comparative Analysis of YOLOv8 and YOLOv11 on Tree Detection Using UAV RGB and Laser Scanning Data

Abstract. To promote sustainable forest management planning including biodiversity monitoring and to enable accurate estimates of stem volume, above-ground biomass, and carbon stocks, tree identification is essential to contemporary forest inventory. Deep learning models are now crucial tools for automating tree recognition over large, forested regions due to the growing availability of high-resolution LiDAR data. In order to identify individual trees using LiDAR-derived RGB raster imagery, this work compares two cutting-edge object identification architectures: YOLOv8 and YOLOv11. A total of 82 annotated images were utilized, rasterized at a resolution of 5 cm, and processed using two input resolutions (640×640 pixel and 960×960 pixel), several model configurations (s, m, l, x), and augmentation settings (rotation and horizontal flip). To provide fair comparison, every model was trained and evaluated using the same methodology. Precision, recall, mAP50, and mAP50-95, standard detection metrics, were used to evaluate performance. The results show that YOLOv8 consistently beat YOLOv11 on all metrics, especially in its large and extra-large forms at high resolution. YOLOv8x with 960 pixel resolution and augmentation was the best-performing setup, with 0.974 precision, 0.837 recall, 0.934 mAP50, and 0.821 mAP50-95. The results demonstrate notable improvements in detection accuracy when compared to previous methods that used YOLOv4 or domain-specific structures like YOLOTree. With the use of rasterized UAV laser scanning data, our results highlight the potential of the YOLO architecture as a robust and scalable tool for automated, high-precision forest inventory.