Independent Hackathon Project • Not an official government warning service

Forecasting Severe-Weather Risk 2–6 Hours Ahead Across West Bengal.

An AI nowcasting model fusing NOAA GFS live atmospheric analyses, trained on four monsoon seasons of ECMWF ERA5 reanalysis, with physics-guided deep learning — estimating thunderstorm, heavy-rain and flash-flood risk across all 23 districts of West Bengal from the Darjeeling Himalayas to the Sundarbans.

Launch StormSense Nowcasting Desk

Live NOAA GFS Ingestion, Refreshed Every 5 Minutes • Calibrated Probabilities, Reported Openly

LOADING CASE STUDY

…

StormSense risk map screenshot from the Cyclone Remal historical case study, showing colour-banded severe-weather risk bands across West Bengal districts

MODEL: SEVEREWEATHERNET V2

GRID: 825 CELLS · 0.25°

LOADING…

PEAK 3H RAINFALL —

Highest District Risk: —

SEVERITY INDEX: — StormSense V2 · Calibrated

FORECAST WINDOW

2–6 Hours

Model lead times from each GFS analysis

LIVE SKILL

~1.8×

PR-AUC vs. no-skill baseline, live GFS input

SPATIAL GRID

~28 km

0.25° grid, 825 cells across 23 districts

REFRESH

5 min

Checks for new GFS cycles (published every 6h)

// MEASURED, NOT CLAIMED

Capabilities: What the Model Actually Does

Two different operating conditions, because they give very different numbers. Lab is held-out ERA5 (the model's own training distribution); live is real NOAA GFS input, which is what the dashboard actually runs on.

HELD-OUT ERA5 TEST (2024 SEASON)

Lab Conditions

Performance on the same reanalysis distribution the model was trained on. This is the best case, not what the live dashboard sees — the live input comes from a different weather system (GFS).

Forecast Leads: +2h to +6h

Spatial Grid: 0.25° (~28 km), 825 cells

CSI (+2h → +6h): 0.394 → 0.244

False Alarm Ratio (FAR): 0.52 – 0.64

LIVE GFS BACKTEST (2024, 14 CASES)

Operational Reality

What the deployed dashboard actually achieves, fed real live NOAA GFS 0.25° analyses. Lower than the lab figures because GFS and ERA5 are genuinely different atmospheric systems.

Forecast Leads: +2h to +6h

Skill vs. No-Skill Baseline: ~1.8× (PR-AUC)

CSI (+2h → +6h): 0.049 → 0.081

False Alarm Ratio (FAR): 0.86 – 0.92

Thunderstorm probability per cell

3-hour rainfall regression

Terrain-aware flash-flood proxy

Per-lead calibrated probabilities

RADAR & SATELLITE FEEDS

Live Radar & Satellite View

A public precipitation radar mosaic (RainViewer), live over the West Bengal domain.

LOADING RADAR

REGIONAL COMPOSITE | SOURCE: RAINVIEWER | COVERAGE: WB + NEIGHBOURS

STORMSENSE MODEL RISK

Normal Watch Alert Warning

DATA SOURCES

RainViewer Radar Mosaic

NOAA GFS 0.25° Analysis

Every 6h

Esri Dark Gray Basemap

Static Tiles

WHAT THIS VIEW SHOWS

Layer	Product	Update	Kind
Radar mosaic	RainViewer composite	~10 min	OBSERVED
Model risk (dashboard)	StormSense severe-weather %	5 min	FORECAST
GFS analysis (input)	NOAA GFS 0.25° f000	6h	OBSERVED

The radar mosaic above is observed precipitation (RainViewer), a different product from the model's own forecast risk surface shown on the dashboard. Open the dashboard for the combined view.

MODEL PIPELINE

How It Works: From Public Weather Data to District Risk

Four stages, from a public global weather analysis to a calibrated, per-district risk estimate.

Refreshes every 5 minutes

01 // Ingest

NOAA GFS Analyses

Surface and pressure-level fields fetched from NOAA's public GFS archive, regridded to the model's domain, and linearly interpolated between the two most recent 6-hourly analyses into six hourly input slots. No forecast hour is ever used as an input.

0.25° Domain, 20–28°N 84–90°E

02 // Harmonise

Match the Training Domain

The model trained on ERA5 but runs live on GFS. Units, sign conventions (e.g. convective inhibition) and precipitation-accumulation windows are aligned so live input means what the training input meant.

ERA5 ↔ GFS Domain Adaptation

03 // Predict

ConvGRU Nowcaster

Separate recurrent encoders read surface, wind and thermodynamic fields; a terrain-conditioned fusion trunk combines them; a shared decoder outputs severe-weather probability and 3-hour rainfall for every lead.

781,889 Parameters

04 // Calibrate & Explain

Physics-Rule Attribution

Probabilities are temperature-calibrated per lead, then combined with a terrain-aware flash-flood proxy. A rule-based physics breakdown (moisture, instability, wind, recent rain) explains each estimate — not a learned attribution method.

Per-Lead Temperature Scaling

INPUT NOAA GFS 0.25° f000 Analyses

COVERAGE 825 Cells, 23 Districts

TRAINING ERA5, May–Oct 2021–2024

STATIC INPUT SRTM Terrain Elevation

// DISTRICT ADVISORIES

District Monitoring: Per-District Risk, Ranked

The dashboard aggregates the 825-cell risk grid into 19 district zones, using the highest-risk cell in each district, and ranks them so the most concerning district always leads.

LOADING LIVE DATA

Loading current district risk…

// IN-APP ALERTING

Impact: What the Alert System Actually Reaches

StormSense can push a real alert to everyone currently viewing the dashboard. It is not connected to any government siren network, SMS gateway, or disaster-response service — the scope below is the honest, current reach.

In-App Siren Broadcast

A full-screen alert with sound and a browser notification, sent to every connected viewer inside the model's current high-risk zone. Location is checked on each viewer's own device and never sent to the server.

Reach: Connected Viewers Only

Advisory Bulletin Log

Informational advisories posted to a standing, readable log — no interruption, no sound. A visitor who opens the dashboard later can still read what was posted.

Persistence: Readable After The Fact

Operational Action Log

An operator can log a request for extended dissemination (e.g. cell broadcast). This does not reach any phone — it is an auditable record for handoff to a real disaster-management authority.

Delivery status: Logged, Not Sent

Honest Scope

No Physical Sirens, No SMS, No Government Integration Today

Reaching people beyond this website — physical sirens, cell broadcast, SMS to the public — requires infrastructure and legal authority owned by government disaster-management bodies and telecom operators. StormSense can supply a risk feed to such a system; it cannot operate one itself.

Reach today: In-App Only

Try the Live Alert System

// ABOUT

About StormSense

An independent hackathon project exploring whether a compact deep-learning model can give useful early signal of severe convective weather over West Bengal from freely available global weather data. Not affiliated with, endorsed by, or operated on behalf of IMD, WBSDMA, NDMA, MoES, or any government body.

Why the Model's Estimate Can Be Checked

Physics-based rule attribution for the current highest-risk cell — not a learned method (no SHAP, no LIME).

LOADING LIVE DATA

Loading current attribution…

Physics-Based, Not a Black Box Every attribution above traces to a real atmospheric field the model consumes — nothing is inferred after the fact.

RULE-BASED ATTRIBUTION

OPEN SOURCE · SELF-HOSTABLE

Want to try it, extend it, or build on it?

The full source, model checkpoints, and every backtest script used to produce the numbers on this page are on GitHub. Run it yourself, verify the results, or fork it for another region.

Open the Dashboard View Source on GitHub Reproduce the Benchmarks

PyTorch FastAPI ECMWF ERA5 NOAA GFS