UrbanTreeFormer: A boundary-aware CNN-transformer architecture for fine-grained urban street tree segmentation

Accurate segmentation of urban street trees from street-level imagery supports tree inventory, ecological assessment, biodiversity monitoring, and smart-city planning. However, irregular crown boundaries, fragmented foliage, occlusion, background clutter, and class imbalance remain significant challenges. This paper presents UrbanTreeFormer, a boundary-aware CNN–Transformer architecture for fine-grained urban tree crown segmentation. The model integrates an EfficientNet-B3 encoder for local feature extraction, a Transformer bottleneck for long-range contextual modeling, an attention-gated decoder for selective feature refinement, and an auxiliary boundary-refinement head to improve crown delineation. A combined weighted cross-entropy, Dice, and boundary loss, alongside class weighting and oversampling, addresses the foreground–background imbalance and the moderate imbalance among species. UrbanTreeFormer was evaluated on the public Urban Street Tree Segmentation dataset comprising 3,949 images and 22 tree species, using binary foreground and species-wise multiclass segmentation protocols. On the 386-image held-out test set, binary segmentation achieved Accuracy of 0.9731 ± 0.0014, Precision of 0.8695 ± 0.0032, Recall of 0.9544 ± 0.0039, Dice of 0.9100 ± 0.0021, and IoU of 0.8349 ± 0.0033. Boundary evaluation yielded BF1 of 0.8810 ± 0.0027, Boundary IoU of 0.7890 ± 0.0040, ASSD of 2.08 ± 0.08 pixels, and HD95 of 6.95 ± 0.21 pixels. Species-wise evaluation achieved macro-averaged Dice and IoU scores of 0.8731 and 0.7778, respectively. UrbanTreeFormer achieved the highest mean performance among eight CNN-, attention-, and Transformer-based baselines, with statistically significant improvements over seven, although its difference from SegFormer was not statistically significant. Ablation studies verified the contribution of each architectural component, while robustness experiments demonstrated stability under moderate photometric variations but reduced performance under severe noise, blur, and compression. These findings highlight UrbanTreeFormer's potential for accurate urban tree crown segmentation and species-level monitoring.