Recent advances and future perspectives in tree crown segmentation for forest monitoring

Tree crown segmentation (TCS) provides crown-level information for forest inventory, biodiversity assessment, biomass estimation and related ecological applications. Its performance, however, remains strongly dependent on canopy structure, sensor characteristics, training data and computational resources. Traditional rule-based methods remain useful in relatively simple stands but often lose accuracy in dense, mixed and vertically layered canopies. Recent advances in unmanned aerial vehicle (UAV) imagery, light detection and ranging (LiDAR), very-high-resolution optical data and learning-based models have improved crown detection and delineation, while multi-sensor approaches combine complementary spectral and structural information. This review synthesises the development of TCS from traditional algorithms to machine learning, deep learning and multi-sensor strategies, with emphasis on their relevance to forest monitoring and downstream ecological applications. The review also assesses major benchmark resources, including the National Ecological Observatory Network (NEON) Tree Crown benchmark, Bamberg Benchmark Forest Dataset of Individual Tree Crowns in Very-High-Resolution UAV Images (BAMFORESTS), Very-High-Resolution Tree Detection (VHRTrees), FOR-instance and OpenAerialMap Tree Crown Delineation (OAM-TCD), and examines differences in sensor coverage, annotation design and ecological representativeness. A structured decision framework is presented to guide method selection according to forest type and canopy complexity, available sensors, labelled-data availability and computing resources. Future priorities include broader benchmark diversity, cross-site and cross-sensor validation, efficient processing and explicit evaluation of how crown-segmentation errors influence downstream ecological indicators.