VHRTrees: a new benchmark dataset for tree detection in satellite imagery and performance evaluation with YOLO-based models

Natural and planted forests, covering approximately 31% of the Earth’s land area, are crucial for global ecosystems, providing essential services such as regulating the water cycle, soil conservation, carbon storage, and biodiversity preservation. However, traditional forest mapping and monitoring methods are often costly and limited in scale, highlighting the need to develop innovative approaches for tree detection that can enhance forest management. In this study, we present a new dataset for tree detection, VHRTrees, derived from very high-resolution RGB satellite images. This dataset includes approximately 26,000 tree boundaries derived from 1,496 image patches of different geographical regions, representing various topographic and climatic conditions. We implemented various object detection algorithms to evaluate the performance of different methods, propose the best experimental configurations, and generate a benchmark analysis for further studies. We conducted our experiments with different variants and hyperparameter settings of the YOLOv5, YOLOv7, YOLOv8, and YOLOv9 models. Results from extensive experiments indicate that, increasing network resolution and batch size led to higher precision and recall in tree detection. YOLOv8m, optimized with Auto, achieved the highest F1-score (0.932) and mean Average Precision (mAP)@0.50 Intersection over Union threshold (0.934), although some other configurations showed higher mAP@0.50:0.95. These findings underscore the effectiveness of You Only Look Once (YOLO)-based object detection algorithms for real-time forest monitoring applications, offering a cost-effective and accurate solution for tree detection using RGB satellite imagery. The VHRTrees dataset, related source codes, and pretrained models are available at https://github.com/RSandAI/VHRTrees.