Commit History
18 days ago
-
Release v2: models trained on the Dynamic World label class
Verified signatureReplaces the August 2026 checkpoint, which was trained on a mixture of binarized Dynamic World water probability (about 65% of tiles) and the DW label band. All models here use the DW label class only, the definition the paper describes. Adds the matched S1-only control and the previous-year AEF model under variants/, the validation summary and threshold sweeps that were missing from v1, and the independent evaluation results on the 53-scene PlanetScope reference set and the S1S2-Water benchmark. src/train.py is now the exact trainer that produced these weights. Pooled water IoU on the 53-scene reference set at threshold 0.30: 0.851 (S1+AEF), 0.840 (previous-year AEF), 0.558 (S1-only), against 0.746 for OPERA DSWx-S1.7c49718