retinanet-tree-groups

A RetinaNet (ResNet-50 FPN) that finds individual trees and tree groups in aerial imagery and draws a box around each one. It was built for HOT's fAIr as a second tree model next to deepforest_tree.

Version v0.1.0: the weights, thresholds and window sizes below are frozen.

Files

File	What it is
best.pt	PyTorch checkpoint (2 classes: 1 tree, 2 tree_group)
model.onnx	The same network exported to ONNX
thresholds.json	Confidence thresholds picked on validation: tree 0.2, tree_group 0.15
windows.json	Sliding-window sizes: 60 m for trees, 120 m for tree groups
config.json	The training configuration

Tree groups that a window edge cuts are dropped (drop_edge), not joined; this rule was picked on validation.

Training

1. Start: DeepForest's NEON.pt tree-crown RetinaNet.
2. Joint base: fine-tuned on NeonTreeEvaluation and OAM-TCD (restor/tcd, restor/tcd-nc), then blended as 30 % of these weights and 70 % of the starting weights.
3. HOT fine-tune: 10 epochs on HOT's Freetown tree and tree_group labels (hotosm/fAIr-guru-td-trees v0.1.0), backbone frozen, random crops of 30-80 m; best epoch 4 on validation.

Imagery is used at zoom 19-21 (about 30-7.5 cm per pixel); validation and test are scored at zoom 20.

Scores

F1 at the validation-picked thresholds, boxes matched at IoU 0.4. The HOT test part (35 units) was held out from training and threshold picking.

Part	Trees F1	Tree groups F1
HOT validation (30 units)	0.573	0.477
HOT test (35 units)	0.531	0.333

Licence and attribution

Weights: CC-BY-NC-4.0. About 6 % of the joint base's OAM-TCD training tiles come from restor/tcd-nc, which is CC-BY-NC-4.0, so the weights carry its non-commercial term.

Attribution:

• HOT tree labels: (c) OpenStreetMap contributors, ODbL-1.0. Imagery: CC-BY-4.0 (hotosm/fAIr-guru-td-trees).
• OAM-TCD: Restor, restor/tcd (CC-BY-4.0) and restor/tcd-nc (CC-BY-NC-4.0).
• NeonTreeEvaluation: Weinstein et al., Zenodo 10.5281/zenodo.5914554 (CC-BY-4.0).
• DeepForest NEON.pt: weecology/deepforest-tree.

Checksums (sha256)

File	sha256
best.pt	cb4cfad93000acddfe1a8b9701901523ffc099baad19583d2a48288368d25fdf
model.onnx	4a59f1c2e958d54cf1bbf9b2ae1083a124043d1ac3a741de391509dc4cce6f8e
thresholds.json	b0b13594471089fc39cf11333a0e2a96990d5e7e29727f7b7e818e70b4c2ca2d
windows.json	e4adbef89e619b3b2a43a46a172edca19eceedd9a952b92a8d4fbc10686f2a7c