A UAV RGB dataset and method for instance tree crown segmentation for biodiversity monitoring

The use of UAVs (Unmanned Aerial Vehicles) equipped with cameras has emerged as a promising approach due to its efficiency and accuracy in assessing biodiversity. Tree instance segmentation plays a crucial role in UAV image analysis, serving as a foundational step for subsequent tasks such as species identification. However, instance segmentation of trees from UAV images presents significant challenges, especially in dense forest environments. This leads to severe ambiguity in determining instance boundaries. Rather than directly segmenting entire tree crowns—which often fails under such challenging conditions—we adopt a deliberate over-segmentation strategy based on a contour detection network. Subsequently, the contours are merged to produce more accurate segmentation results. These ideas are integrated into the first contribution of the paper: TreeCoG - a method for instance-level tree crown segmentation from UAV-captured RGB images. TreeCoG consists of three main steps: (1) contour extraction, (2) contour feature extraction and representation, and (3) contour merging. The first step extracts contour candidates from UAV images, while the second and the third steps construct a graph from these contours and employ a Graph Convolutional Network (GCN) to merge them into instances. The second contribution of this paper is a new image dataset, ForestSeg, collected using UAVs in dense tropical forests of Vietnam in multiple seasons. The ForestSeg dataset is composed of four subsets—ForestSeg-T1, ForestSeg-T2, ForestSeg-T3, and ForestSeg-T4 comprising a total of 2,944 annotated images. These subsets were acquired at different time periods and flight altitudes, thereby capturing variations in tree appearance and supporting robust evaluation of instance tree crown segmentation methods. This dataset enables the analysis of temporal variations in tree characteristics and serves as a valuable resource for evaluating the robustness of segmentation methods under varying tree appearances over time. We conduct extensive experiments to assess the contribution of each component of TreeCoG and compare our approach with the state-of-the-art methods on two datasets: our self-collected dataset, ForestSeg, and the BAMFORESTS dataset. Experimental results demonstrate the superior performance of TreeCoG in tree instance segmentation, achieving 57.01% AP, 62.21% AP@50, and 55.32% AP@70 on the ForestSeg dataset, and 53.21% AP, 73.14% AP@50, and 43.23% AP@70 on the BAMFORESTS dataset, respectively. These results confirm the effectiveness of the proposed method in accurately delineating individual tree instances.