AIML based Nowcasting of thunderstorm and lightning using atmospheric observation including multiple radars, satellite, lightning and model data.
Ministry of Earth Sciences (MoES) · Disaster Management · Software
Free satellite imagery makes convective nowcasting genuinely achievable and the verification methodology lets you prove real skill against a fair baseline — but the lightning data the title asks for is not public, so scope to convection, beat advection, and be honest about the gap.
What it actually is
Thunderstorms build and dissipate within a couple of hours, which is far too fast for a normal weather forecast to be useful, and lightning kills a large number of people in India every year. The ask is a system that predicts where a storm and its lightning will be over the next hour or two from radar, satellite and lightning observations as they arrive.
What to build
A nowcasting model producing zero-to-two-hour forecasts of convective activity over a defined region on a fixed grid, learning storm initiation, growth, decay and motion from a sequence of satellite infrared and water-vapour imagery and radar reflectivity where accessible rather than only advecting the current field forward, outputting a gridded probability of thunderstorm activity per time step with a separate lightning-likelihood layer where observation data supports it, evaluated with the categorical scores nowcasting is properly judged by — probability of detection, false alarm ratio and critical success index at each lead time — against both a persistence baseline and an optical-flow advection baseline that is genuinely hard to beat.
Smallest thing that wins the room
Run the model on a held-out convective afternoon and play the two-hour forecast loop beside what actually happened, with the critical success index against the advection baseline shown for each lead time.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 44% of the 226 · #127 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: central ministry statements sat below the average.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5Nowcasting is genuinely doable from free satellite imagery, the verification methodology is well established so you can prove skill against a fair baseline, and the safety case for lightning warning in India is unarguable — but the lightning observations the statement names are not openly available, so you will be nowcasting convection and inferring lightning rather than doing what was asked.
Feasibility
3/5Geostationary satellite infrared and water-vapour imagery over India is freely available at high cadence and is enough to nowcast convective development on its own, but the radar products are only selectively public and bulk lightning location data for India is not openly distributed, so the lightning half of the statement rests on observations you probably cannot obtain.
Innovation scope
4/5The description is the title repeated, prescribing no architecture, region, grid, lead time or output format, so the entire formulation of the nowcasting problem is yours.
Clarity
1/5A verbatim repetition of the title with no background, no requirement, no lead time, no resolution, no data source and no accuracy criterion — the statement provides no information about the deliverable beyond naming the task.
Effort
HeavyBuilding an aligned multi-source image sequence archive, training a spatiotemporal model, implementing meaningful baselines and constructing a proper categorical verification harness is four pieces, with the data alignment across sources being the slowest.
Demo-ability
MediumA forecast loop playing beside the observed sequence is a good and honest visual, but convective nowcasting output is a probability field rather than a crisp answer, and its quality is only apparent to someone who understands what the baseline would have produced.
In its favour
- Green flag: Geostationary satellite imagery over India is free and frequent enough to nowcast convection on its own, so the project is viable even without radar or lightning access
- Green flag: Nowcasting has a well-established verification methodology with standard categorical scores, so you can demonstrate skill rigorously rather than showing a plausible-looking animation
- Green flag: Optical flow advection is a genuinely strong and fair baseline, and beating it convincingly is a real result that a meteorological panel will recognise immediately
- Green flag: Convective initiation — predicting a storm that does not yet exist — is the part advection fundamentally cannot do, and it is where a learned model has a legitimate and demonstrable advantage
Against it
- Red flag: Bulk lightning location data for India is not openly distributed, so the lightning half of the statement cannot be trained or verified properly and you will be inferring it from convective proxies — say so rather than implying you predicted lightning
- Red flag: Advection of the current radar or satellite field is a surprisingly strong nowcasting baseline at short lead times, and a model that merely matches it has added nothing
- Red flag: Convective events are rare in a grid-and-time sense, so a model optimised on accuracy will learn to predict no storm and score well — use categorical scores that penalise misses, not accuracy
- Red flag: The description contains nothing but the title, so two teams could build entirely different systems and both be responsive, which puts the burden of defining success on you
What you will be writing
- ConvLSTM or U-Net spatiotemporal nowcasting
- INSAT-3D IR and water vapour channel sequences
- optical flow advection baseline
- POD, FAR and CSI verification by lead time
- convective initiation from brightness temperature cooling rate
- gridded probability output with lead-time animation
- Convective nowcasting
- Satellite meteorology
- Severe weather warning
Prior art to read before you start
short-range convective nowcasting · spatiotemporal sequence prediction from imagery · categorical forecast verification against baselines
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.