Dhara flood segmentation model (unet_r34_imagenet_v2)

Flood segmentation model used by Dhara, which maps flood and debris damage and cut-off communities from Sentinel-1 radar. Built for the Multimodal AI Hackathon 2026, Track B. Educational prototype, not for operational decisions.

Model

Architecture U-Net, ResNet-34 encoder (ImageNet-pretrained)
Input 6 × 224 × 224, Sentinel-1 σ⁰ in dB: post VV, post VH, pre₁ VV, pre₁ VH, pre₂ VV, pre₂ VH, standardised with norm.json
Output 3 classes: no water, permanent water, flood
Calibration Temperature 1.3964 (in model_card.json)
Format ONNX, runs on CPU with ONNX Runtime

Training data

Kuro Siwo (Bountos et al., NeurIPS 2024), CC BY. Four high-relief "mountain" events were held out of training, chosen by DEM relief before training (splits.json).

Results

Evaluation Score
Kuro Siwo test events, flood F1 0.74
Held-out mountain events, flood F1 0.75
Sen1Floods11 test, water IoU 0.61
Sen1Floods11 Bolivia (held-out country), water IoU 0.71

Details: metrics.json, metrics_sen1floods11.json, and the validation notes.

Use

Place model.onnx in models/flood_unet/ of the Dhara repository; model_card.json and norm.json are already there. SHA-256 of model.onnx: 038c6d80eab8a40639f2c64c5ab79036548b7eb5719d913dd3fd8e1d8a13005a

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support