Transformer-Based Individual Tree Crown Detection from Canopy Height Models with Cross-Domain and Self-Supervised Pretraining

Individual tree crown (ITC) detection from remotely sensed data is fundamental to forest inventory and ecological monitoring, but deep learning approaches remain constrained by limited labelled training data. We systematically evaluate three transformer detectors (the Detection Transformer (DETR), Deformable DETR, and DETR with Improved DeNoising Anchor Boxes (DINO)) paired with two backbones, ImageNet-pretrained ResNet-50 and a Masked Autoencoder (MAE) pretrained on unlabelled Canopy Height Model (CHM) data. These are benchmarked against a classical local maximum and watershed pipeline and Faster R-CNN across four test sets spanning boreal, temperate mixed-wood, and diverse North American forest types at 0.25–1.0 m resolution. Spatially held-out test regions with a one-patch buffer band eliminate sliding-window leakage; headline configurations are reported as mean ± standard deviation across three random seeds. With multi-resolution MAE pretraining, the practical lower bound for non-degenerate single-dataset transformer detection lies between ∼200 and ∼1200 patches. Without MAE pretraining, DETR fails at every dataset size we tested. Multi-dataset joint training reaches F1=0.84±0.01 on the boreal test set and 0.45–0.68 across the temperate-mixed-wood and NEON test sets, while Faster R-CNN narrowly wins on the smallest training pool. Standard DETR with ResNet-50 collapses regardless of the length of training schedule, but the same architecture with an MAE backbone reaches F1=0.83±0.01 at that schedule, showing that DETR’s reported instability is conditional on the combination of backbone initialization and training budget rather than architectural. Resolution and backbone interact: ResNet-50 wins at 0.25 m, and MAE wins at 0.5–1.0 m, consistent with the eight-pixel MAE patch-matching crown scale only at coarser resolutions.