Eight stages run on every new satellite scan (every 30 minutes for INSAT-3D). The highlighted stage follows the replay above, so you can see where each number on the dashboard comes from.

AI engine

Three models, one forecast

Each task gets a model suited to its data: a detector for finding the storm in an image, a regressor for its strength, and a sequence model for where it goes next.

Identification

Cyclone detection

YOLOv8 + U-Net cloud segmentation

Input: IR1 10.8 µm and WV 6.7 µm brightness temperature
Finds the circulation centre and cloud shield extent
Labels from IMD / IBTrACS best-track positions
Output: bounding box, centre lat/lon, confidence

Classification

Intensity estimation

ResNet-50 CNN regression + softmax head

Learns Dvorak-style features: eye, CDO, banding
Predicts max wind (km/h) and central pressure
Maps to 7 IMD classes, D through SuCS
Fused with ASCAT surface winds when available

Prediction

Track forecasting

ConvLSTM + ERA5 steering flow

Past 24 h of positions, motion and intensity
Environmental fields: 500 hPa winds, SST, shear
Predicts centre positions at +6, +12, +18 and +24 h
Ensemble spread gives the uncertainty cone

Database · Backend / API

Where results are stored and served

PostgreSQL with PostGIS keeps every storm fix, prediction and alert. The backend (PHP or Node.js Express on your own server) serves them as JSON.

Tables

PostgreSQL + PostGIS

Table	Holds	Key fields
storms	One row per system	id, name, basin, genesis_time
fixes	Observed centre every 6 h	storm_id, time, geom, wind, pressure
predictions	AI track + intensity	storm_id, issued, lead_h, geom, wind
detections	Satellite AI results	image_time, prob, t_number, category
alerts	Risk engine output	location, level, eta_h, issued

API endpoints

JSON over HTTPS

Method	Endpoint	Returns
GET	/api/live	Active storms, latest track, official forecast, IMD warnings
GET	/api/sat	Latest INSAT-3DR IR1 image (proxied)
GET	/api/predict?storm=	24 h track prediction (planned)
GET	/api/alerts	Location risk levels (planned)