← All experiments
PROTOTYPE · IN PROGRESS
Automatic Deforestation Detection
Prototype U-Net model trained on high-resolution satellite imagery to detect deforestation. Includes a 44-epoch training curve inspector and an adjustable confidence cutoff preview.
Threshold Preview
128x128 tile
SYNTHETIC FIELD · p >= 0.90
Non-deforestation
Deforestation
0.0 Low
0.5 Mid
1.0 High
0.90
Deforestation area
0.0%
Default view is a synthetic probability field. Local PNG uploads threshold grayscale brightness directly without running in-browser model inference.
Training Run (44 Epochs)
Train
Validation
LR: 1e-6 (EarlyStop)
Epoch 44 / 44
Train: 0.844
Val: 0.731
Val P / R: 0.93 / 0.89
Final test (42 tiles): 0.460 IoU · 99.1% recall · 69.6% precision
Open Work
-
[ ]
Export sample 128x128 test imagery and model probability outputs into
examples/to replace the synthetic preview field. - [ ] Evaluate quantized ONNX export for client-side inference.