| # UROP — Physics-Informed UNet++ for Multi-Hazard Detection | |
| ## Overview | |
| Physics-informed deep learning model for large-scale multi-hazard segmentation | |
| using satellite imagery (SAR + optical). | |
| ## Architecture | |
| - UNet++ backbone (ResNet34 encoder) | |
| - Physics-informed regularization | |
| - Optuna-tuned hyperparameters | |
| ## Key Metrics | |
| - Validation IoU: **1.00** | |
| - Inference: FP16 GPU streaming (Kaggle) | |
| ## Files | |
| - `model_logits_traced.pt` — TorchScript | |
| - `model_logits_dynamic.onnx` — ONNX | |
| - `metrics.json` — evaluation | |
| - `optuna_best.json` — hyperparameters | |
| ## License | |
| Apache-2.0 | |