From what to build to how to build it — our platform helps you understand, plan, and solve SIH 2026 problem statements with complete insights into requirements, solution approaches, technology stacks, architecture, AI/ML, and implementation roadmaps.
Search for your problem statement, and start building your solution.
| Description: | AIML based Nowcasting of thunderstorm and lightning using atmospheric observation including multiple radars, satellite, lightning and model data. |
| Organization: | Ministry of Earth Sciences (MoES) |
| Department: | India Meteorological Department |
| Category: | Software |
| Theme: | Disaster Managementtd> |
| Youtube Link: | Not provided |
| Dataset Link: | Not provided |
| Contact info: | Not provided |
Thunderstorms and lightning represent severe atmospheric hazards in India, causing over 2,000 casualties annually and causing extensive destruction to livestock, aviation, power grids, and agricultural infrastructure [KNOWN]. Traditional Numerical Weather Prediction (NWP) models require multi-hour computational cycles and fail to resolve the micro-scale convective initiation and rapid decay of localized thunderstorm cells within 0–3 hour lead times [RESEARCHED]. This dossier presents a production-grade, AI/ML-driven multimodal nowcasting architecture engineered specifically for the Ministry of Earth Sciences (MoES) - India Meteorological Department (IMD) under Problem Statement SIH26072 [PS]. The proposed solution fuses high-frequency Doppler Weather Radar (DWR) reflectivity grids, INSAT-3D/3DR/3DS geostationary satellite multispectral imagery, Indian Institute of Tropical Meteorology (IITM) Lightning Location Network (LLN) stroke events, and WRF atmospheric instability fields into a hybrid Deep Spatiotemporal Vision Transformer with Physics-Guided Advection [RESEARCHED]. The resulting platform delivers 1 km × 1 km hyper-local thunderstorm reflectivity (dBZ) and lightning stroke probability forecasts updated every 15 minutes, with an automated Common Alerting Protocol (CAP) alerting engine [INFERENCE].
| Field | Value | Confidence |
|---|---|---|
| Problem Statement ID | SIH26072 |
[PS] |
| Title | AIML based Nowcasting of thunderstorm and lightning using atmospheric observation including multiple radars, satellite, lightning and model data. | [PS] |
| Organization | Ministry of Earth Sciences (MoES) | [PS] |
| Department | India Meteorological Department (IMD) | [PS] |
| Category | Software | [PS] |
| Theme | Disaster Management | [PS] |
| Deadline | 30 September 2026 | [PS] |
| Ideas Counter | 0/500 | [PS] |
| Fundamental Objective | Build a real-time, multi-modal AI/ML pipeline and interactive GIS decision support dashboard that ingest radar, satellite, lightning sensor, and NWP model data to predict thunderstorm cell tracking and lightning hazards 0–3 hours ahead at 1 km resolution with sub-minute inference latency. | [INFERENCE] |
The explicit problem statement provided by MoES/IMD requires an AI/ML-based nowcasting system utilizing four distinct atmospheric observation streams [PS]. Below is the complete atomic deconstruction of explicit requirements and derived technical implications.
REQ-EX-01: AI/ML-Based Nowcasting System
[PS][INFERENCE].[INFERENCE].[PS]REQ-EX-02: Severe Weather Focus (Thunderstorm and Lightning)
[PS][INFERENCE].[INFERENCE].[PS]REQ-EX-03: Multi-Radar Data Integration
[PS][INFERENCE].[INFERENCE].[PS]REQ-EX-04: Satellite Data Integration
[PS][INFERENCE].[RESEARCHED].[PS]REQ-EX-05: Ground Lightning Sensor Data Integration
[PS][INFERENCE].[RESEARCHED]. Conversion of point stroke events into continuous spatial density rasters [INFERENCE].[PS]REQ-EX-06: NWP Model Data Integration
[PS][INFERENCE].[RESEARCHED].[PS]REQ-IN-01: Real-Time Ingestion & Low-Latency Inference ([INFERENCE])
[INFERENCE].REQ-IN-02: Hyper-Local Spatial Resolution ([INFERENCE])
[INFERENCE].REQ-IN-03: GIS Web & Mobile Decision Support Dashboard ([INFERENCE])
[INFERENCE].| Stakeholder Group | Goal | Current Pain Point | System Need |
|---|---|---|---|
| IMD Operational Meteorologists | Issue precise 0–3h nowcasts and district/station-level thunderstorm warnings. | Manual inspection of disparate radar, satellite, and lightning windows; high cognitive load during rapid convective development [RESEARCHED]. |
Unified GIS interface fusing all observational feeds with AI-generated storm vectors and automated warning polygon suggestions [INFERENCE]. |
| National / State Disaster Management Authorities (NDMA / SDMA) | Evacuate open fields, halt outdoor operations, and deploy emergency teams before lightning strikes. | Coarse district-wide alerts with multi-hour lag lead to warning fatigue and missed hyper-local severe strikes [KNOWN]. |
Automated CAP-compliant alerts triggered by 15-minute lead-time lightning probability thresholds at $1\text{ km}^2$ grid level [INFERENCE]. |
| Farmers & Agricultural Workers | Take shelter during open-field farming when a lightning storm approaches within 10–20 km. | Lack of hyper-local, low-bandwidth real-time alerts in regional languages leading to high mortality rates [KNOWN]. |
Low-latency mobile notification / SMS alert with regional audio-visual warnings and 30-minute lead time [INFERENCE]. |
| Aviation & Airport Authorities (AAI / ATC) | Re-route flights and halt tarmac ground handling during severe convective activity and wind shear. | Delayed radar-only updates fail to capture rapid cloud-top electrification and microburst initiation [KNOWN]. |
Automated runway cell velocity vectors and 0–2h reflectivity corridor extrapolation [INFERENCE]. |
| Power Distribution Companies (DISCOMs) | Isolate vulnerable transmission grids and dispatch repair crews for storm-damaged lines. | Unpredictable lightning strikes trigger widespread grid trips without localized pre-warning [KNOWN]. |
Sub-station-level lightning risk heatmaps and strike count prediction [INFERENCE]. |
| Proposed System Feature | Justifying Requirement | Core Technical Implementation | Confidence |
|---|---|---|---|
| Multi-Modal Feature Ingestion Engine | REQ-EX-03, REQ-EX-04, REQ-EX-05, REQ-EX-06 | Async Python pipeline using netCDF4, pyart, and rasterio to parse DWR NetCDF/BUFR, INSAT-3D HDF5, IITM LLN CSV, and WRF GRIB2 files [RESEARCHED]. |
[INFERENCE] |
| Spatio-Temporal Grid Resampling & Mosaicing | REQ-EX-03, REQ-EX-04 | Equal-area WGS84 EPSG:4326 grid interpolation engine normalizing multi-sensor resolutions to a standard $1\text{ km} \times 1\text{ km}$ spatial tensor [RESEARCHED]. |
[INFERENCE] |
| Cross-Attention Vision Transformer & U-Net Model | REQ-EX-01, REQ-EX-02 | Deep learning neural network ingesting past $T-4$ to $T_0$ multi-channel frames (Radar Z, INSAT CTT/WV, LLN Density, WRF CAPE) to forecast future $T_{+1}$ to $T_{+12}$ frames (0–3h) [RESEARCHED]. |
[INFERENCE] |
| Dual Prediction Heads (Reflectivity + Lightning) | REQ-EX-02 | Parallel output layers predicting 1) Max Reflectivity (dBZ) raster and 2) Calibrated Lightning Stroke Probability map $P(\text{Stroke} > 0) \in [0, 1]$ [INFERENCE]. |
[INFERENCE] |
| WebGIS Interactive Warning Dashboard | REQ-IN-03 | Mapbox GL JS / Leaflet frontend rendering time-series radar opacity overlays, satellite cloud motion, lightning stroke history, and CAP warning polygons [ASSUMPTION]. |
[INFERENCE] |
| Automated CAP Alerting & Push Gateway | REQ-IN-01, REQ-IN-02 | Rules engine triggering XML/JSON CAP alerts via WebSockets, Firebase Cloud Messaging (FCM), and SMS when cell reflectivity $> 40\text{ dBZ}$ and lightning probability $> 70%$ [RESEARCHED]. |
[INFERENCE] |
Nowcasting systems in India operate under severe environmental, infrastructural, and observational constraints [KNOWN].
┌─────────────────────────────────────────────────────────┐
│ INDIAN OPERATIONAL ENVIRONMENT REALITIES │
└────────────────────────────┬────────────────────────────┘
│
┌───────────────────────┬───────────────────┴───────────────────┬───────────────────────┐
│ │ │ │
┌────────┴────────┐ ┌────────┴────────┐ ┌────────┴────────┐ ┌────────┴────────┐
│ RADAR COVERAGE │ │ SATELLITE LATENCY│ │ BANDWIDTH REALITY│ │ FIELD POWER │
│ GAPS & TERRAIN │ │ & RESOLUTION │ │ IN RURAL AREAS │ │ & INFRASTRUCTURE│
└────────┬────────┘ └────────┬────────┘ └────────┬────────┘ └────────┬────────┘
│ │ │ │
┌───────┴────────┐ ┌───────┴────────┐ ┌───────┴────────┐ ┌───────┴────────┐
│• Non-uniform │ │• 15-30 min scan│ │• 2G/3G networks│ │• Frequent power│
│ radar mesh │ │ interval │ │ in farming │ │ outages during│
│• Terrain beam │ │• 4 km spatial │ │ districts │ │ thunderstorms │
│ blockage in │ │ resolution │ │• 100 KB max │ │• Edgeware sync │
│ Himalayas │ │ (TIR channels)│ │ payload limit │ │ resilience │
└────────────────┘ └────────────────┘ └────────────────┘ └────────────────┘
[RESEARCHED]. Coverage is dense along the east coast and major metros but sparse in parts of Central India, the Northeast, and mountainous Himalayan regions where radar beam blockage occurs [RESEARCHED]. The AI model must gracefully fallback to satellite and NWP inputs in radar-blind zones [INFERENCE].[RESEARCHED]. Temporal alignment and spatial feature resampling are major processing bottlenecks [INFERENCE].[KNOWN]. Heavy map tiles cannot be transmitted; the platform must deliver lightweight vector polygons ($< 50\text{ KB}$) or plain SMS alerts [INFERENCE].[INFERENCE].To ensure technical rigor, the following critical research questions must be addressed during development:
[RESEARCHED].[RESEARCHED].[RESEARCHED].[RESEARCHED].| Data Type | Purpose | Ideal Real-World Source | Hackathon Substitute | Access Method | Availability | Privacy Sensitivity | Confidence |
|---|---|---|---|---|---|---|---|
| Doppler Weather Radar (DWR) Reflectivity | Convective storm structure, intensity, and advection tracking. | IMD Doppler Weather Radar Network (37+ stations, Max Z / VIL products) [RESEARCHED]. |
NOAA NEXRAD Level-II reflectivity grids or open radar datasets on AWS Open Data [RESEARCHED]. |
S3 / HTTP download / PyART NetCDF parsing [RESEARCHED]. |
Public archive available; Real-time API restricted [NEEDS VALIDATION]. |
Non-sensitive (Public weather) | [RESEARCHED] |
| Geostationary Satellite Infrared / WV | Cloud top cooling rates, convective cloud initiation outside radar coverage. | ISRO/MOSDAC INSAT-3D / 3DR / 3DS Imager & Sounder (TIR1, TIR2, WV, CTT) [RESEARCHED]. |
GOES-16/17 Advanced Baseline Imager (ABI) or EUMETSAT MSG SEVIRI open netCDF archives [RESEARCHED]. |
MOSDAC API / FTP or AWS Open Data Cloud Storage [RESEARCHED]. |
Public archival access via MOSDAC registration [RESEARCHED]. |
Non-sensitive | [RESEARCHED] |
| Ground Lightning Stroke Events | Ground truth targets for lightning stroke density and flash probability heads. | IITM Pune Lightning Location Network (83+ sensors) / Damini App backend [RESEARCHED]. |
ENGLN (Earth Networks) public samples, WWLLN (World Wide Lightning Location Network) or GOES Geostationary Lightning Mapper (GLM) [RESEARCHED]. |
CSV / GeoJSON point streams or gridded netCDF [RESEARCHED]. |
Sample datasets public; Live feed requires MoES MOU [NEEDS VALIDATION]. |
Non-sensitive | [RESEARCHED] |
| NWP Atmospheric Thermodynamics | Large-scale background instability (CAPE, CIN, Shear, Lifted Index). | IMD Operational WRF (3 km) / NCUM model outputs [RESEARCHED]. |
NOAA Global Forecast System (GFS) 0.25° or ERA5 Reanalysis atmospheric variables [RESEARCHED]. |
NOMADS GRIB2 feed or Copernicus CDS API [RESEARCHED]. |
Publicly available free global access [KNOWN]. |
Non-sensitive | [KNOWN] |
| Digital Elevation Model (DEM) | Terrain slope/elevation impact on orographic convective uplift. | SRTM 90m / ISRO CartoDEM [RESEARCHED]. |
AWS Terrain Tiles / USGS SRTM 30m / 90m [RESEARCHED]. |
Cloud-Optimized GeoTIFF (COG) via HTTP [KNOWN]. |
Publicly available free global access [KNOWN]. |
Non-sensitive | [KNOWN] |
EXISTING APPROACHES vs PROPOSED SOLUTION
┌───────────────────────────────┐ ┌───────────────────────────────┐ ┌───────────────────────────────┐
│ Extrapolation Models │ │ Standard Physics NWP │ │ PROPOSED HYBRID AI/ML │
│ (Optical Flow / TITAN) │ │ (WRF / NCUM) │ │ MULTIMODAL FUSION │
├───────────────────────────────┤ ├───────────────────────────────┤ ├───────────────────────────────┤
│ • Purely advective motion │ │ • High physical accuracy │ │ • Fuses Radar, INSAT, LLN & │
│ • Zero cell growth/decay skill│ │ • 2-4 hour compute delay │ │ WRF parameters │
│ • Blurs quickly beyond 45 mins│ │ • Poor micro-scale resolution │ │ • 15-second GPU inference │
│ • Blind to initial convection │ │ • High false alarm rate 0-1h │ │ • Models growth & decay │
└───────────────────────────────┘ └───────────────────────────────┘ └───────────────────────────────┘
[RESEARCHED].[RESEARCHED].[RESEARCHED].[INFERENCE].[RESEARCHED].[KNOWN].[RESEARCHED].[INFERENCE].[RESEARCHED].[RESEARCHED].[RESEARCHED].[INFERENCE].[RESEARCHED].[RESEARCHED].[INFERENCE].[INFERENCE]. SOLUTION APPROACH COMPARISON
┌─────────────────────────────────┬─────────────────────────────────┬─────────────────────────────────┐
│ APPROACH 1: CLASSICAL │ APPROACH 2: SPATIOTEMPORAL │ APPROACH 3: MULTIMODAL VIZ │
│ FEATURE + XGBOOST │ 2D ConvLSTM / U-NET │ TRANSFORMER + OPTICAL FLOW │
├─────────────────────────────────┼─────────────────────────────────┼─────────────────────────────────┤
│ • Tabular grid features │ • Image-to-image sequence model │ • Dual-branch hybrid arch │
│ • Fast, interpretable │ • Learns motion implicitly │ • Optical flow advection core │
│ • Ignores spatial context │ • Blurs at longer lead times │ • Cross-attention for physics │
│ • Low hackathon complexity │ • Moderate GPU overhead │ • Sharp, physically grounded │
└─────────────────────────────────┴─────────────────────────────────┴─────────────────────────────────┘
[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[RESEARCHED].[RESEARCHED].[RESEARCHED].[INFERENCE].[RESEARCHED].[RESEARCHED].[INFERENCE].[RESEARCHED].[INFERENCE].[INFERENCE].The Hybrid Physics-Aligned Vision Transformer + Optical Flow Advection (Approach 3) is selected as the winning architecture [INFERENCE].
[INFERENCE].[RESEARCHED].[RESEARCHED]. This approach achieves high Critical Success Index (CSI) scores while maintaining sharp operational warning boundaries [INFERENCE].====================================================================================================
1. DATA INGESTION & ADAPTER LAYER (Async Ingestion Workers)
├─ IMD DWR Feed (NetCDF/BUFR) ──► PyART Parser ──────┐
├─ MOSDAC INSAT-3D/3DR (HDF5) ──► GDAL/SatPy Parser ├──► Spatial Grid Resampler & Aligner
├─ IITM LLN Feed (CSV/WebSocket)► Point-to-Raster │ (WGS84 EPSG:4326, 1 km Mesh)
└─ Operational WRF (GRIB2) ────► xarray Extractor ──┘
====================================================================================================
2. FEATURE PREPROCESSING & TENSOR BUILDER
├─ Normalization & Missing Value Imputation (Spatial Kriging)
├─ Multi-Channel Stacking: [Radar Z, VIL, INSAT TIR1, WV, CTT, LLN Density, CAPE, Shear]
└─ Rolling Spatiotemporal Memory Buffer (T-60m, T-45m, T-30m, T-15m, T0)
====================================================================================================
3. HYBRID AI/ML NOWCASTING INFERENCE ENGINE
├─ Branch A: Neural Optical Flow Advection Network (Predicts Motion Vectors u, v)
├─ Branch B: Cross-Attention Vision Transformer (Ingests WRF CAPE/Shear + INSAT CTT)
├─ Feature Fusion & Advective Warping Layer
└─ Parallel Prediction Heads:
├─ Head 1: Reflectivity Forecast Grid [T+15m ... T+180m] (dBZ Raster)
└─ Head 2: Lightning Stroke Probability Map [T+15m ... T+180m] (P in [0,1])
====================================================================================================
4. GEOSPATIAL POST-PROCESSING & WARNING ENGINE
├─ Polygon Extraction: Marching Squares Contour Algorithm (Z >= 40 dBZ, P_light >= 0.7)
├─ District / Station Spatial Overlay Intersect Engine (PostGIS)
└─ Common Alerting Protocol (CAP v1.2) XML/JSON Alert Generator
====================================================================================================
5. PRESENTATION & DELIVERY LAYER
├─ FastAPI REST API & WebSocket Gateway
├─ Interactive WebGIS Dashboard (Mapbox GL JS / React / Deck.gl)
├─ Mobile FCM / WhatsApp / SMS Alert Notification Service
└─ Low-Bandwidth GeoJSON Vector Tile Server
====================================================================================================
[INFERENCE]. DWR reflectivity volumes are parsed via PyART, INSAT-3D HDF5 multispectral bands via SatPy, IITM LLN point stroke logs via Pandas/Geopandas, and WRF GRIB2 files via xarray [RESEARCHED].[RESEARCHED]. The aligned arrays are normalized and stacked into a 5D PyTorch tensor [INFERENCE].[RESEARCHED]. Simultaneously, the Cross-Attention ViT calculates growth/decay factors based on satellite cloud-top temperature drops and WRF atmospheric instability [RESEARCHED].[RESEARCHED]. Hazard polygons ($> 40\text{ dBZ}$ / $> 70%$ lightning risk) are intersected with administrative boundaries (districts, sub-districts, airport buffers) in PostGIS [INFERENCE].[INFERENCE].| Layer | Selected Technology | Why Selected | Alternative Considered | Why Alternative Rejected |
|---|---|---|---|---|
| Frontend UI | React + TypeScript + Mapbox GL JS / Deck.gl | High-performance WebGL rendering of multi-layered spatial raster overlays and vector polygons at 60 FPS [KNOWN]. |
Leaflet.js | Struggles with rendering high-density time-series animated raster tiles and 10,000+ lightning vector markers [KNOWN]. |
| Backend API | FastAPI (Python 3.11) + Uvicorn | Native async I/O, auto-generated OpenAPI docs, seamless integration with PyTorch AI runtime [KNOWN]. |
Node.js / Express | Lacks native Python scientific libraries (numpy, xarray, torch, pyart) required for raster data processing [KNOWN]. |
| Geospatial Database | PostgreSQL 16 + PostGIS 3.4 | Industry-standard spatial indexing (GIST), fast polygon intersection queries, and native raster functions [KNOWN]. |
MongoDB / GeoJSON | Inferior spatial joining performance and lacks specialized GIS projections and spatial functions [KNOWN]. |
| AI/ML Framework | PyTorch 2.3 + CUDA 12.1 + TorchScript | Superior support for spatiotemporal Vision Transformers, dynamic tensor shapes, and TensorRT acceleration [KNOWN]. |
TensorFlow / Keras | Slower innovation cycle for state-of-the-art vision models; complex multi-modal dynamic graph execution [KNOWN]. |
| Geospatial Processing | PyART + SatPy + Rasterio + xarray | Native support for weather radar formats (NetCDF/BUFR), satellite HDF5 channels, and multi-dimensional arrays [RESEARCHED]. |
GDAL CLI Scripts | Heavy shell wrappers add execution overhead and lack native Python memory buffers [INFERENCE]. |
| Time-Series / Caching | Redis 7.0 | Ultra-fast in-memory caching for current T0 spatial tensors, active WebSocket sessions, and rate-limiting [KNOWN]. |
Memcached | Lacks geospatial pub/sub capabilities and complex data structures (hashes, sorted sets) [KNOWN]. |
| Task Queue / Message Broker | Celery + RabbitMQ | Reliable asynchronous execution of heavy file parsing, raster interpolation, and model inference jobs [KNOWN]. |
Apache Kafka | Overkill for hackathon scale; introduces higher infrastructure setup complexity for student team [INFERENCE]. |
| Containerization | Docker + NVIDIA Container Toolkit | Ensures exact GPU driver compatibility, library dependency isolation, and reproducible deployments [KNOWN]. |
Bare-metal execution | High risk of "works on my machine" CUDA version mismatches during live hackathon judging [INFERENCE]. |
AI/ML MODEL ARCHITECTURE
┌─────────────────────────────────────────────────────────────────────────────────────────┐
│ MULTIMODAL INPUT TENSOR │
│ [ Radar Z (t-4..t0) | INSAT TIR1/WV (t-4..t0) | LLN Density (t-4..t0) | WRF CAPE/Shear ]│
└────────────────────────────┬────────────────────────────────┬───────────────────────────┘
│ │
▼ ▼
┌───────────────────────────────────────────┐ ┌───────────────────────────────────────────┐
│ NEURAL OPTICAL FLOW BRANCH │ │ CROSS-ATTENTION VISION TRANSFORMER │
│ • Computes motion field (u, v) │ │ • Ingests thermodynamic background │
│ • Performs backward advective warping │ │ • Predicts cell growth/decay residual ΔZ │
│ • Preserves sharp reflectivity edges │ │ • Predicts electrification potential │
└────────────────────────────┬──────────────┘ └────────────────────────────┬──────────────┘
│ │
└────────────────┬───────────────┘
│
▼
┌───────────────────────────┐
│ FEATURE FUSION MODULE │
└─────────────┬─────────────┘
│
▼
┌─────────────────────────────────┐
│ DUAL PREDICTION HEADS │
├─────────────────────────────────┤
│ Head 1: Reflectivity Forecast │
│ Head 2: Lightning Probability │
└─────────────────────────────────┘
[RESEARCHED].[RESEARCHED].[INFERENCE].[RESEARCHED].[RESEARCHED].[RESEARCHED].[RESEARCHED].[INFERENCE].[INFERENCE].[RESEARCHED].[RESEARCHED].[RESEARCHED].[RESEARCHED].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE]. AI is NOT used for immediate strike echoing, preserving sub-second alert response [INFERENCE].-- Active Weather Radar Observations Grid Table
CREATE TABLE radar_observations (
obs_id BIGSERIAL PRIMARY KEY,
station_id VARCHAR(10) NOT NULL,
observation_time TIMESTAMPTZ NOT NULL,
max_reflectivity_raster RASTER NOT NULL,
created_at TIMESTAMPTZ DEFAULT NOW()
);
CREATE INDEX idx_radar_obs_time ON radar_observations(observation_time);
-- Lightning Stroke Point Logs (IITM Network Ingestion)
CREATE TABLE lightning_strokes (
stroke_id BIGSERIAL PRIMARY KEY,
stroke_time TIMESTAMPTZ NOT NULL,
location GEOMETRY(Point, 4326) NOT NULL,
peak_current_ka REAL,
stroke_type VARCHAR(5) CHECK (stroke_type IN ('CG', 'IC'))
);
CREATE INDEX idx_lightning_geom ON lightning_strokes USING GIST(location);
CREATE INDEX idx_lightning_time ON lightning_strokes(stroke_time);
-- AI Generated Forecast Polygons
CREATE TABLE forecast_hazard_polygons (
polygon_id BIGSERIAL PRIMARY KEY,
issue_time TIMESTAMPTZ NOT NULL,
valid_time TIMESTAMPTZ NOT NULL,
lead_time_minutes INT NOT NULL,
hazard_type VARCHAR(20) CHECK (hazard_type IN ('THUNDERSTORM', 'LIGHTNING', 'HAIL')),
risk_level VARCHAR(10) CHECK (risk_level IN ('YELLOW', 'ORANGE', 'RED')),
geom GEOMETRY(Polygon, 4326) NOT NULL,
confidence_score REAL CHECK (confidence_score BETWEEN 0 AND 1)
);
CREATE INDEX idx_hazard_poly_geom ON forecast_hazard_polygons USING GIST(geom);
.pbf format) [KNOWN]. File size per tile is reduced by $> 85%$ ($< 15\text{ KB}$ per tile) [INFERENCE].[KNOWN]. When connectivity drops, users can still view the last downloaded 3-hour forecast trend and local emergency guidelines [INFERENCE].[INFERENCE].[KNOWN]. The ingestion pipeline implements spatial masking algorithms over designated defense coordinates [INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[KNOWN]. To prevent panic caused by unvalidated automated warnings, public outputs must clearly state: "Experimental AI-Assisted Forecast - Powered by IMD/MoES Nowcasting Engine" [INFERENCE].[INFERENCE]. SIH HACKATHON MVP SCOPE
┌──────────────────────────────────────────────────┐ ┌──────────────────────────────────────────────────┐
│ IN SCOPE (MUST BUILD) │ │ DESCRIBED ONLY / OUT OF SCOPE │
├──────────────────────────────────────────────────┤ ├──────────────────────────────────────────────────┤
│ • Regional Target Zone (e.g., East India Grid) │ │ • Pan-India 37+ DWR live streaming hardware sync │
│ • 4-Channel Ingestion (Radar, INSAT, LLN, WRF) │ │ • Direct CAP integration with telecom operators │
│ • PyTorch Hybrid AI Nowcast Engine (T+0 to T+2h) │ │ • Hardware-level Doppler radar signal processing │
│ • Interactive WebGIS Dashboard (Mapbox GL JS) │ │ • Retraining pipelines on multi-terabyte archives│
│ • CAP Alert Generator & WhatsApp/FCM Simulator │ │ • 24/7 high-availability multi-region cloud cluster│
└──────────────────────────────────────────────────┘ └──────────────────────────────────────────────────┘
[ASSUMPTION].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].To deliver a compelling live demonstration during SIH judging without relying on live government network connections, the team will implement a transparent, realistic simulation pipeline [INFERENCE].
LIVE DEMO SIMULATION ARCHITECTURE
┌────────────────────────────────┐ ┌────────────────────────────────┐
│ HISTORICAL WEATHER EVENT │ │ SIMULATED REAL-TIME INGEST │
│ REPLAY ENGINE │ │ STREAMER │
│ • Kalboishakhi Storm Event │ │ • WebSocket push every 10s │
│ • Pre-processed NetCDF/HDF5 │────────────────►│ • Simulates T0, T-15, T-30 │
│ • Real historical strikes │ │ • Passes frames to AI engine │
└────────────────────────────────┘ └───────────────┬────────────────┘
│
▼
┌────────────────────────────────┐ ┌────────────────────────────────┐
│ LIVE MOBILE & WEBGIS │ │ AI MODEL INFERENCE & CAP │
│ ALERT RECEIVER │ │ GENERATOR │
│ • Displays live animated maps │◄────────────────│ • Runs PyTorch model live │
│ • Triggers FCM push notification│ │ • Generates alert polygons │
└────────────────────────────────┘ └────────────────────────────────┘
[INFERENCE].demo_streamer.py) that reads pre-processed historical radar, satellite, and lightning slices, pushing them into the Redis input queue every 10 seconds to simulate a live 15-minute operational ingestion loop [INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].Sized specifically against the final deadline of 30 September 2026 [PS]. Assuming a 4-week active development window leading up to final submission:
IMPLEMENTATION ROADMAP
┌──────────────────────┬──────────────────────┬──────────────────────┬──────────────────────┐
│ WEEK 1: DATA & BASE │ WEEK 2: AI MODEL │ WEEK 3: BACKEND & │ WEEK 4: GIS DASHBOARD│
│ INGESTION PIPELINE │ TRAINING & TUNING │ GEOSPATIAL ENGINE │ INTEGRATION & DEMO │
├──────────────────────┼──────────────────────┼──────────────────────┼──────────────────────┤
│ • Parse NetCDF/HDF5 │ • Train PyTorch ViT │ • PostGIS spatial │ • Build Mapbox UI │
│ • Spatial resampling │ advection model │ polygon engine │ • Mobile FCM alerts │
│ • Grid alignment │ • Evaluate CSI/FSS │ • Fast-API endpoints │ • Live demo replay │
│ • Synthetic dataset │ • Optimize CUDA │ • CAP XML generator │ • Final testing & │
│ preparation │ inference latency │ • Redis queue setup │ documentation │
└──────────────────────┴──────────────────────┴──────────────────────┴──────────────────────┘
[INFERENCE].rasterio and scipy.interpolate to project all inputs onto a unified $1\text{ km} \times 1\text{ km}$ grid [INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].[INFERENCE].POD > 0.80) [INFERENCE].[RESEARCHED].[INFERENCE].| Risk Event | Likelihood | Impact | Severity | Mitigation Strategy | Confidence |
|---|---|---|---|---|---|
| Real-time IMD DWR API access unavailable during hackathon | High | High | HIGH | Use NOAA NEXRAD open radar grids or pre-downloaded MOSDAC/IMD historical NetCDF archives [INFERENCE]. |
[ASSUMPTION] |
| Temporal misalignment between INSAT (30m) and Radar (10m) | High | Medium | MEDIUM | Implement temporal linear feature interpolation and forward-propagation frame alignment in Redis buffer [INFERENCE]. |
[INFERENCE] |
| Deep learning model causes blurry forecasts at T+2 hours | Medium | High | HIGH | Decouple advective motion via Neural Optical Flow and apply Perceptual/Focal Loss instead of pure MSE [RESEARCHED]. |
[RESEARCHED] |
| GPU memory overflow during multi-channel tensor processing | Medium | Medium | MEDIUM | Crop target domain to $256 \times 256$ or $512 \times 512$ grid tiles and utilize FP16 mixed precision inference [INFERENCE]. |
[INFERENCE] |
| Public alert panic due to false positive lightning predictions | Low | High | HIGH | Implement forecaster approval gate and require confidence score threshold $> 0.75$ before CAP alert dispatch [INFERENCE]. |
[INFERENCE] |
[KNOWN].$0) [INFERENCE].[INFERENCE].[RESEARCHED].[INFERENCE].[INFERENCE].[INFERENCE]. KEY SYSTEM DIFFERENTIATORS
┌─────────────────────────────────┬─────────────────────────────────┬─────────────────────────────────┐
│ FEATURE FUSION ENGINE │ HYBRID OPTICAL FLOW + ViT │ CAP & LOW-BANDWIDTH │
├─────────────────────────────────┼─────────────────────────────────┼─────────────────────────────────┤
│ Combines 4 distinct feeds │ Eliminates spatial blurring │ Generates standardized CAP │
│ (Radar + Satellite + Lightning │ maintaining crisp 1 km │ v1.2 alerts & vector tiles │
│ + NWP) into a single tensor │ warning polygons up to 3 hours │ for 2G network mobile delivery │
└─────────────────────────────────┴─────────────────────────────────┴─────────────────────────────────┘
[INFERENCE].[RESEARCHED].[INFERENCE].[RESEARCHED].[INFERENCE].[INFERENCE].[ASSUMPTION][ASSUMPTION-01] Regional Target Domain: The hackathon MVP will target a $500\text{ km} \times 500\text{ km}$ regional grid over Eastern India (Odisha/West Bengal) rather than full national pan-India coverage to fit GPU hardware limits.[ASSUMPTION-02] Radar Data Format: Radar reflectivity files are assumed to be accessible in standard NetCDF4 or IRIS BUFR formats compatible with PyART.[ASSUMPTION-03] Demo Streaming: A historical replay streamer is accepted by judges as a valid demonstration mechanism for real-time atmospheric data flows.[NEEDS VALIDATION][NEEDS VALIDATION-01] Direct Real-Time API Access: Availability of unauthenticated live REST/FTP endpoints for IMD DWR and IITM Damini feeds during the hackathon window.[NEEDS VALIDATION-02] Cloud Top Temperature Thresholds: Exact operational CTT drop rate thresholds ($\Delta\text{CTT}/\Delta t$) used by IMD meteorologists for pre-radar convective initiation validation.[INFERENCE].[INFERENCE].[INFERENCE].[RESEARCHED].[INFERENCE].[RESEARCHED].rasterio, satpy, and pyart to project satellite and radar rasters onto a standard $1\text{ km}$ WGS84 grid [INFERENCE].[INFERENCE].[INFERENCE].demo_streamer.py to push sequential pre-processed frames into Redis for live demo testing [INFERENCE].PyART NetCDF loading [INFERENCE].88/100 (Highly feasible for a skilled 4–6 student team utilizing open Python geospatial and deep learning libraries) [INFERENCE].[INFERENCE].[RESEARCHED][PS, KNOWN].[RESEARCHED][RESEARCHED].[RESEARCHED][RESEARCHED].[RESEARCHED][RESEARCHED].[RESEARCHED][RESEARCHED].[RESEARCHED][RESEARCHED].What we think about this platform, and how it can help people solve SIH problems.
" 30M+ Google Gemini tokens went into researching, analyzing, and understanding all 233 SIH problem statements to make them easier to explore and solve. "
" The platform makes SIH problem statements much easier to understand. Having everything from the problem details to the procedure in one place. "
" Oh, you got ads? Well, we need to make money too! But, you can use our platform to help you win SIH and take home the prize money. "
Get in touch if you need help bringing your ideas to life, or simply have an interesting idea you'd like to talk through. I'm always open to a conversation and exploring new possibilities together.