Moriae/ukraine-forest-change-detection

Forest Change Detection Models

Staging folder for uploading the main individual model artifacts to Hugging Face.

Training Dataset

These models were trained on the Ukraine forest change detection dataset:

https://huggingface.co/datasets/Moriae/ukraine-forest-change-detection

Code Repository

Model definitions, training code, evaluation scripts, and ONNX export utilities are available here:

https://github.com/Moriae1337/satellite-forest-change-detection

Evaluation Results

Performance on the Test Set

Model	Threshold	Accuracy	Precision	Recall	F1-Score	IoU	Kappa
Siamese FCNN	0.35	0.9885	0.7997	0.6821	0.7362	0.5826	0.7304
Attention U-Net	0.40	0.9895	0.7952	0.7464	0.7700	0.6260	0.7646
Attention Siamese U-Net	0.40	0.9897	0.8170	0.7273	0.7696	0.6254	0.7643
ChangeFormer V6	0.40	0.9891	0.7895	0.7173	0.7557	0.6074	0.7502
ChangeFormer V6+	0.45	0.9887	0.7753	0.7332	0.7537	0.6047	0.7479
ChangeMamba BCD	0.45	0.9889	0.8099	0.6944	0.7477	0.5971	0.7421

Legend: Bold indicates the best result (1st), and Italics indicate the second-best result (2nd).

Included checkpoints

• checkpoints/unet/best_model.pth
• checkpoints/siamese_unet/best_model.pth
• checkpoints/siamese_fcnn/best_model.pth
• checkpoints/changeformer/best_model.pth
• checkpoints/changemamba_bcd/best_model.pth
• checkpoints/changeformer_plus/best_model.pth

Included ONNX exports

• onnx/unet/unet_logits.onnx
• onnx/unet/unet_logits.onnx.data
• onnx/siamese_fcnn/siamese_fcnn_logits.onnx
• onnx/siamese_unet/siamese_unet_logits.onnx
• onnx/changeformer/changeformer_logits.onnx
• onnx/changeformer_plus/changeformer_plus_logits.onnx
• onnx/changeformer_plus/changeformer_plus_logits.onnx.data
• onnx/changemamba_bcd/changemamba_bcd_logits.onnx

The .onnx.data files are required external tensor data for the matching .onnx files.

Metadata

• config.json contains the model, dataset, artifact, and inference defaults used by this project.
• metrics/history_summary.csv summarizes training histories for the individual runs.
• metrics/*_best_threshold.json contains selected validation/test thresholds where available.