Deep learning-based automatic tree-level crown detection from high-resolution UAV imagery for aboveground carbon...

Accurate delineation of individual tree crowns is essential for forest resource management and carbon stock estimation. However, traditional manual interpretation is inefficient under complex forest conditions, and existing deep learning-based object detection models remain vulnerable to crown overlap and background interference. To address these limitations, this study proposes an improved YOLOv8-CAFM model. By integrating the local feature extraction capability of convolutional neural networks (CNNs) with the global contextual modeling capability of transformers, the CAFM module enhances the detection of small and partially occluded crowns. Two plantation forests with contrasting canopy densities, dominated by Pinus sylvestris var. mongolica and Pinus tabuliformis in Zhangwu County, Liaoning Province, China, were selected as the study sites. An individual tree crown dataset was manually delineated from high-resolution UAV imagery to support automated crown detection. The proposed model was compared with widely used crown detection models, including Faster R-CNN, YOLOx, and DeepForest. Based on the detected crowns, field measurements, species-specific allometric equations, and carbon content coefficients were further combined to estimate individual-tree aboveground carbon (AGC) stocks. The results showed that YOLOv8-CAFM performed well in both sparse and dense forests, achieving mAP@0.5:0.95 values of 0.736 and 0.609, respectively, and F1-scores of 0.974 and 0.856, outperforming Faster R-CNN, YOLOx, and DeepForest. The extracted crown areas were also highly consistent with the reference data (R² = 0.898 for sparse forests and 0.794 for dense forests), demonstrating strong boundary delineation capability. Furthermore, the estimated forest AGC stocks showed significant correlations with field measurements, with R² values of 0.53 for Pinus sylvestris var. mongolica and 0.60 for Pinus tabuliformis. By jointly modeling local and global features, the proposed method improves individual tree detection, particularly in dense-canopy stand, and provides efficient technical support for accurate forest carbon sink monitoring.