Model Weights

The .pth weights and corresponding model hyperparameters in .yaml files for models described in the publication are provided.

S2 Multispectral Flood Mapping Model

• s2_unet_all_best.pth - benchmarked UNet model using all channels (except DEM).
  • Use threshold t=0.5 for inference. Achieves 89.77% F1, 91.69% Recall, 87.94% Precision on test set.
• s2_unetpp_all_best.pth - benchmarked UNet++ model using all channels (except DEM).
  • Use calibrated threshold t=0.75 for inference. Achieves 89.98% F1, 91.12% Recall, and 88.86% Precision on test set.

S1 SAR Flood Mapping Model

• s1_unetpp_all_best.pth - benchmarked UNet++ SAR model using all channels (except DEM).
  • Use calibrated threshold t=0.85 for inference. Achieves 75.32% F1, 63.61% Recall, and 92.35% Precision on test set.
• s1_unetpp_all_cvae_best.pth - benchmarked UNet++ SAR model using all channels (except DEM), specifically trained for CVAE despeckled SAR images.
  • Use calibrated threshold t=0.87 for inference. Achieves 76.06% F1, 64.82% Recall, and 92.02% Precision on test set.

S1 Despeckling Model

• cvae.pth - benchmarked SAR CVAE despeckler for VV and VH channels with 25.92-27.09 PSNR, 0.63-0.66 SSIM and 188.7-379.5 ENL on test set.