Review of deep learning for tree species identification from high-resolution remote sensing imagery

Zhongyu Xia , Teja Kattenborn, Jan Dirk Wegner, Arthur Gessler, Verena C. Griess, Mirela Beloiu

Tree species information is essential for monitoring forest composition, carbon storage, biodiversity, and resilience to environmental change, but remains largely available only at local scales. Centimeter-resolution remotely sensed imagery and deep learning are shifting species mapping from stands to individual tree crowns, yet the diversity of tree species, data, and methods has fragmented this rapidly growing field. Here, we provide a systematic review to assess progress and define research priorities for crown-level tree species identification based primarily on centimeter-resolution 2D imagery and deep learning. We synthesized 133 studies published between 2017 and 2025, representing 1,034 deep learning-based crown-level species identification tasks across diverse ecosystems, sensor types, and modeling approaches. We found strong geographic and taxonomic biases, with 83% of studies concentrated on temperate and subtropical forests. RGB was the dominant input (62% of tasks), followed by multispectral (10%) and hyperspectral imagery (9%), while 17% used multimodal data. Among controlled 2D–3D comparisons, 60% showed improved performance. Among computer vision tasks, patch-wise classification accounted for 41%, followed by semantic segmentation (29%), object detection (18%), and instance segmentation (13%). CNNs remained the predominant architecture (77%) but new transformer-based and hybrid architectures are emerging. Cross-site transferability remains the central challenge to closing geographic and taxonomic gaps, requiring representative open datasets and standardized evaluation across sites and sensors.