drishti-flood-nowcaster

Drishti Flood Nowcaster — Baseline U-Net (SIH 26085)

Baseline urban-flood nowcasting model for Smart India Hackathon 26085 (Urban Flood Nowcasting System, MoES/NCMRWF): 30-min rainfall/depth history + static terrain maps → flood depth at 9 lead times (5–180 min) on a 110×160 @5 m grid.

⚠️ Trained on SYNTHETIC data. All drainage networks are terrain-inferred (verified=false), never surveyed KIET infrastructure. Flood depths are physics-model output, never observed flooding. Demo/research use only — not for engineering decisions.

Files

• pytorch_model.bin — main weights (100-scenario train, 1.9 MB; {"model": state_dict, "in_ch": 36})
• baseline_subset30.pt — earlier 30-scenario smoke-run weights
• baseline_unet.py — architecture (BaselineUNet, 476k params)
• training.log — full training log

Metrics (validation, 49 scenarios / 542 windows)

Loss = MSE(depth) + 0.2·BCE(flooded mask ≥ 5 cm). Test/OOD eval pending (streaming loader).

Usage

import torch
from baseline_unet import BaselineUNet

m = BaselineUNet(in_ch=36, n_leads=9)
m.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu")["model"])
m.eval()
# x: (B, 36, 110, 160) = 9 static + 18 dyn-hist + 9 future-rain (train-normed)
# y: (B, 9, 110, 160) flood depth in metres at leads [5,10,20,30,40,60,90,120,180] min
with torch.no_grad():
    y = m(torch.zeros(1, 36, 110, 160))

Training data

Stratified synthetic scenarios (rainfall class quotas + Latin-Hypercube sampling): 227 train / 49 val / 48 test scenarios + 34 OOD. Physics: diffusive-wave storage-cell solver (Bates & De Roo 2000) + Manning pipes, mass conserved ~1e-8. Normalization (normalization_train.json) computed from TRAIN only.

Source

Built in the Drishti repo — see report.md §§16–18, docs/dataset_report.md, knowledge.md §6.1.