Adaptive multi-source instance segmentation for individual tree crown delineation using UAV multispectral imagery and canopy height models

Accurate delineation of individual tree crowns is essential for precision forest management, carbon accounting, and ecological monitoring, yet it remains challenging due to crown overlap, shadowing effects, and the limited ability of single-source imagery to jointly capture canopy spectral and structural information. Although satellite remote sensing provides broad spatiotemporal coverage, the spatial resolution of freely available imagery is generally insufficient for reliable individual tree-level mapping. Unmanned Aerial Vehicles (UAVs) can acquire high-resolution multi-source remote sensing data from local to regional scales; however, reliance on isolated data sources still constrains the joint characterization of canopy spectral variability and vertical structure. To address these challenges, we propose an adaptive multi-source instance segmentation framework that integrates UAV multispectral imagery (R, G, B, NIR, and RE bands) with canopy height models (CHMs). Built upon Mask R-CNN, the framework incorporates a Confidence Map Estimator (CME) to improve boundary reliability and an Adaptive Fusion Module (AFM) to balance cross-modal feature contributions at the instance level. In addition, edge-based feature selection, together with cross-occupancy instance merging, is introduced to mitigate boundary artifacts induced by image tiling and to enhance global segmentation consistency. Experimental results demonstrate that the proposed framework consistently outperforms the RGB-only baseline in individual tree crown segmentation. The optimal configuration achieves a mean Average Precision (mAP) of 0.678 and a mean Average Recall (mAR) of 0.755, corresponding to improvements of 10.6 and 6.6 percentage points, respectively. For forest parameter estimation, canopy closure prediction shows an R 2 improvement of 0.11 and a 6.12% reduction in relative RMSE, while individual tree counts result in an R 2 increase of 0.07 and a 4.85% reduction in relative RMSE. Overall, this study presents an integrated workflow encompassing UAV-based multi-source data acquisition, adaptive cross-modal fusion, instance segmentation, and crown delineation. The results demonstrate that synergistic multi-source integration with adaptive fusion improves the accuracy and consistency of individual tree crown delineation and stand-level forest parameter estimation, providing a scalable and deployable solution for forest management, carbon accounting, and related ecological monitoring.