Problem Statements Analysis

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.

Scroll down
01 / SIH 2026

PS: SIH26074

Downscaling of weather forecast from Block level to Panchayat level: Inferring high-resolution plots/ data/ information from low-resolution plot /data /information /variables for agro-meteorological advisory services.

Description: Downscaling of weather forecast from Block level to Panchayat level: Inferring high-resolution plots/ data/ information from low-resolution plot /data /information /variables for agro-meteorological advisory services.
Organization: Ministry of Earth Sciences (MoES)
Department: India Meteorological Department
Category: Software
Theme: Agriculture, FoodTech & Rural Developmenttd>
Youtube Link: Not provided
Dataset Link: Not provided
Contact info: Not provided

1. Executive Snapshot

The India Meteorological Department (IMD) under the Ministry of Earth Sciences (MoES) currently generates medium-range weather forecasts at the Block level (covering ~6,500 blocks across India) under the Gramin Krishi Mausam Sewa (GKMS) scheme [RESEARCHED]. However, micro-climatic variations caused by local topography, elevation gradients, land-use patterns, water body proximity, and canopy cover cause significant weather discrepancies across Gram Panchayats (~2,60,000 across India) within the same block [INFERENCE]. This dossier provides a complete technical, data, and architectural strategy for building a Physics-Informed Deep Learning & Geostatistical Weather Downscaling Engine that ingests low-resolution Block-level forecasts (~10–25 km grid), fuses high-resolution spatial covariates (30m DEM elevation, land cover, vegetation indices, distance to water bodies), and produces hyper-local Panchayat-level (~1–3 km spatial grid) weather predictions and stage-specific Agromet Advisories [INFERENCE].

Metadata Table

Field Detail
PS Number SIH26074 [PS]
Title Downscaling of weather forecast from Block level to Panchayat level: Inferring high-resolution plots/ data/ information from low-resolution plot /data /information /variables for agro-meteorological advisory services [PS]
Organization Ministry of Earth Sciences (MoES) [PS]
Department India Meteorological Department [PS]
Category Software [PS]
Theme Agriculture, FoodTech & Rural Development [PS]
Deadline 30 September 2026 [PS]
Ideas Counter 0/500 [PS]

Fundamental System Objective: To build an automated, physics-guided AI downscaling framework and localized Agromet decision-support platform that transforms coarse block-level weather forecasts into hyper-local Panchayat-level micro-forecasts and actionable crop-wise farm advisories [INFERENCE].

2. Problem Deconstruction

The problem statement explicitly requires inferring high-resolution weather variables from low-resolution block forecasts specifically for agrometeorological advisories [PS].

Atomic Requirements Breakdown

  • R1: Block-to-Panchayat Spatial Downscaling [PS]

    • Source Text: "Downscaling of weather forecast from Block level to Panchayat level" [PS]
    • Interpretation: The system must accept weather forecast variables (temperature, precipitation, relative humidity, wind speed, cloud cover) provided at the administrative block resolution (~10–25 km) and output spatial grids or point estimates corresponding to individual Gram Panchayats (~1–3 km) [INFERENCE].
    • Implementation Implication: Requires spatial disaggregation algorithms, digital elevation model (DEM) integration, lapse-rate corrections, and spatial interpolation or deep super-resolution neural networks [INFERENCE].
    • Confidence: [PS]

  • R2: High-Resolution Information Inference from Low-Resolution Inputs [PS]

    • Source Text: "Inferring high-resolution plots/ data/ information from low-resolution plot /data /information /variables" [PS]
    • Interpretation: The model must utilize fine-grained environmental covariates (elevation, slope, aspect, land surface temperature, land use/land cover, NDVI) as auxiliary inputs to reconstruct sub-grid spatial variability that is omitted in coarse Numerical Weather Prediction (NWP) outputs [INFERENCE].
    • Implementation Implication: Machine learning / deep learning super-resolution pipeline (e.g., U-Net / SRCNN / Gradient Boosting) trained on historical gridded observations (e.g., IMD 0.25° grid, ERA5-Land, CHIRPS) combined with high-resolution GIS layers [INFERENCE].
    • Confidence: [PS]

  • R3: Agro-Meteorological Advisory Integration [PS]

    • Source Text: "...for agro-meteorological advisory services." [PS]
    • Interpretation: The downscaled weather parameters must not remain raw numerical values; they must feed into an agrometeorological rules engine to generate crop-specific, growth-stage-specific, and weather-impacted advisory recommendations (e.g., spray windows, irrigation scheduling, frost alerts, heat stress mitigation) [INFERENCE].
    • Implementation Implication: Development of a rule-based and AI-assisted Agromet Advisory Service (AAS) engine mapped to ICAR/IMD crop-weather threshold databases (e.g., KALP/SANKALP frameworks) [RESEARCHED].
    • Confidence: [PS]

3. Stakeholders & User Needs

Stakeholder Goal Current Pain System Need
Smallholder Farmers (Panchayat Level) Receive accurate weather warnings and farm advisories specific to their village/farm plot [INFERENCE]. Block forecast fails to capture micro-climates (e.g., valley frost, hill rain shadows), leading to wasted pesticide/irrigation inputs [RESEARCHED]. Hyper-local Panchayat weather forecasts, voice/multilingual alerts, clear actionable advice (e.g., "Do not spray pesticide today in Panchayat X due to 80% rain chance") [INFERENCE].
District Agromet Units (DAMU) / AMFU Scientists Produce tailored agromet advisory bulletins twice weekly (Tuesday/Friday) efficiently [RESEARCHED]. Manually scaling block forecasts down to local villages is time-prohibitive and subjective [INFERENCE]. Automated spatial downscaling platform with pre-calculated Panchayat weather grids and draft advisory suggestions for scientific validation [INFERENCE].
Agricultural Extension Workers (Krishi Sakhis, KVKs) Disseminate verified advisory alerts to last-mile rural communities [RESEARCHED]. Generic block bulletins are ignored by farmers because they don't match local field realities [INFERENCE]. Panchayat-filtered dashboard, WhatsApp/SMS push triggers, and offline-capable mobile app for field visits [INFERENCE].
IMD / MoES Scientists & Administrators Expand Gramin Krishi Mausam Sewa (GKMS) coverage from 6,500 blocks to 2,60,000 Gram Panchayats accurately [RESEARCHED]. Physical AWS/ARG weather stations cannot be installed in every Panchayat due to extreme capital and maintenance costs [INFERENCE]. Software-driven AI downscaling pipeline that maximizes ROI on existing AWS networks and satellite remote sensing data [INFERENCE].

4. Requirements Traceability

Proposed Feature Target Requirement Justification & Technical Bridge
Terrain-Aware Physics-Guided Super-Resolution Neural Network (U-Net + DEM Covariates) R1 & R2 [PS] Fuses coarse block forecasts with high-resolution SRTM 30m DEM, Slope, Aspect, and Sentinel/MODIS NDVI layers to infer sub-grid temperature and rainfall distributions [INFERENCE].
Gram Panchayat Geofencing & Boundary Mapper (LGD Code Integration) R1 [PS] Maps spatial downscaled raster grids directly onto official Local Government Directory (LGD) Panchayat vector boundaries [RESEARCHED].
Rule-Based Agromet Threshold Decision Engine (Crop-Stage Matrix) R3 [PS] Converts downscaled variables (e.g., 2m Max Temp > 38°C during wheat flowering stage) into immediate actionable mitigation advisories [RESEARCHED].
Offline-First PWA & WhatsApp Bot for Last-Mile Dissemination R3 [PS] Ensures smallholder farmers in low-connectivity Gram Panchayats receive timely alerts in regional languages [INFERENCE].
Spatial Validation & Error Heatmap Dashboard for AMFU Scientists R1 & R3 [PS] Provides confidence scores, RMSE metrics, and human-in-the-loop override capability for domain experts before public release [INFERENCE].

5. Constraints & Environment Realities

  • Geographic & Topographic Heterogeneity: India's agro-climatic zones range from Western Ghats micro-valleys to arid Rajasthan plains [KNOWN]. Flat terrain displays low spatial variance, whereas mountainous or coastal terrain displays high micro-climate variance [INFERENCE].
  • Sparse Observational Ground Truth: Automatic Weather Stations (AWS) and Automatic Rain Gauges (ARGs) are distributed unevenly, often 20–50 km apart [RESEARCHED]. Ground truth for every single Panchayat does NOT exist [INFERENCE].
  • Rural Connectivity Limitations: 2G/3G networks dominate deep rural Gram Panchayats [KNOWN]. Offline caching, low-bandwidth vector maps, and minimal payload size (<50 KB per request) are mandatory [INFERENCE].
  • Digital Literacy & Vernacular Diversity: Farmers prefer voice notes, infographics, or simple color-coded alerts over complex weather parameter graphs [RESEARCHED].
  • Computational Constraints: Real-time dynamical downscaling (e.g., running high-resolution WRF model at 1 km across India) requires petascale supercomputing (e.g., NCMRWF Arunika) [RESEARCHED]. The hackathon and production system MUST use fast ML/statistical downscaling that runs inference in seconds on standard cloud VMs [INFERENCE].

6. Research Questions

  1. What is the baseline spatial resolution of current operational IMD block-level forecasts under GKMS?
    • Answer: Block forecasts are currently generated at ~12 km to 25 km spatial resolution twice weekly (5-day outlook) [RESEARCHED].
  2. What physics/environmental covariates most strongly govern spatial weather variations at 1–3 km scale?
    • Answer: For Temperature: Elevation (Lapse Rate ~6.5°C/km), Aspect, Canopy Cover/NDVI, Land Surface Temperature (LST) [RESEARCHED]. For Precipitation: Orographic uplift (slope/elevation), distance to coast/water bodies, Radar reflectivity, convective cloud top temperature [RESEARCHED].
  3. What is the official administrative structure for Gram Panchayat spatial boundaries in India?
    • Answer: The Ministry of Panchayati Raj maintains the Local Government Directory (LGD) coding system and spatial GeoJSON/shapefiles via Bharat Maps and Bhuvan ISRO portals [RESEARCHED].
  4. How can deep learning models perform spatial downscaling without ground-truth weather stations in every Panchayat?
    • Answer: By training on high-resolution reanalysis datasets (ERA5-Land at 9 km, CHIRPS at 0.05° ~5 km, IMD 0.25° gridded precipitation, NASA POWER) coupled with high-resolution static DEM/LULC layers using Super-Resolution Convolutional Neural Networks (SRCNN/U-Net) or Physics-Informed Neural Networks [RESEARCHED].
  5. What operational downscaling systems does MoES/IMD currently pilot?
    • Answer: IMD piloted meteoGAN (300m Delhi-NCR rainfall downscaling) and a 1 km AI-enabled rainfall forecast pilot for Uttar Pradesh in May 2026 [RESEARCHED].

7. Data Requirements & Sources

Data Categories Breakdown

A. Data Explicitly Mentioned in Problem Statement

  • Low-resolution Block-level weather forecast parameters (Temperature, Rainfall, Humidity, Wind) [PS].
  • Panchayat-level downscaled output variables [PS].

B. Data Required by Inference

  • Digital Elevation Model (DEM) data (30m SRTM / CartoDEM) for elevation, slope, aspect [INFERENCE].
  • Land Use / Land Cover (LULC) and Normalized Difference Vegetation Index (NDVI) [INFERENCE].
  • Gram Panchayat administrative boundaries with unique LGD codes [INFERENCE].
  • Crop-weather phenology calendars and crop stress threshold matrices [INFERENCE].

C. Data That Must Be Collected / Ingested Operationalizing System

  • Real-time block weather forecasts from IMD API / Open-Meteo / ECMWF [INFERENCE].
  • Daily AWS / ARG observational weather records for model validation and bias correction [INFERENCE].

D. Data That Can Be Synthetically Generated (Hackathon Phase)

  • Historical multi-year block forecast vs Panchayat station observations for model fine-tuning [ASSUMPTION].
  • Mock farmer registration database with land holding coordinates and active crop stages [ASSUMPTION].

E. Data That Still Needs Validation

  • Real-time access permissions for IMD's internal operational block forecast REST API [NEEDS VALIDATION].

Data Source Matrix

Data Type Purpose Ideal Real-World Source Hackathon Substitute Access Method Availability Privacy Sensitivity Confidence
Block Weather Forecast Primary low-res input (5-day outlook) IMD GKMS / Agromet API [RESEARCHED] Open-Meteo API / GFS 0.25° forecast [KNOWN] REST API / NetCDF / GRIB2 Public / Open Public Domain [RESEARCHED]
Elevation & Terrain (DEM) Spatial downscaling covariate (Lapse rate, slope) ISRO CartoDEM (10m) / NASA SRTM 30m [RESEARCHED] USGS SRTM 30m GeoTIFF via Earth Engine [KNOWN] Cloud Optimized GeoTIFF (COG) Public Public Domain [KNOWN]
Vegetation & Land Cover Surface roughness & evapotranspiration features ISRO Bhuvan LULC / Copernicus Sentinel-2 NDVI [RESEARCHED] MODIS NDVI (250m) / ESA WorldCover (10m) [KNOWN] Google Earth Engine / SpatioTemporal Asset Catalog (STAC) Public Public Domain [KNOWN]
Panchayat Boundaries Vector overlay for spatial aggregation Survey of India / Ministry of Panchayati Raj (LGD) [RESEARCHED] OpenCity / Community GeoJSON Panchayati Raj shapefiles [KNOWN] GeoJSON / PostGIS Vector Tiles Open Community Public Domain [KNOWN]
Historical Weather Grids Downscaling ML model training target IMD 0.25° Rainfall & 0.5° Temp Grids / ERA5-Land [RESEARCHED] ERA5-Land (9km) daily reanalysis (2018–2025) [RESEARCHED] Copernicus Climate Data Store (CDS) API / Zarr Open Public Domain [RESEARCHED]
Crop-Weather Thresholds Agromet advisory generation engine ICAR-CRIDA Agromet Domain Rules / KALP platform [RESEARCHED] Standard ICAR agronomic handbook rules matrix [RESEARCHED] JSON/SQL Rulebook Database Semi-Public Public Domain [RESEARCHED]

8. Existing Solutions & Gap Analysis

Existing System Description Strengths Weaknesses Relevance Remaining Gap Source Confidence
IMD GKMS / Meghdoot App Official MoES mobile app providing district/block weather forecasts and twice-weekly AAS bulletins [RESEARCHED]. Authoritative, covers 6,500 blocks, vernacular support in 27 languages [RESEARCHED]. Forecasts limited to Block level; cannot capture village-level topographic micro-climates [RESEARCHED]. Direct baseline system that our solution enhances [INFERENCE]. Lacks automated spatial downscaling to Panchayat level (~1–3 km grid) [INFERENCE]. PIB / IMD [RESEARCHED] [RESEARCHED]
IMD KALP Platform Location-specific weather-based crop advisory tool developed by IMD Agrimet [RESEARCHED]. Dynamic comparison of forecasts against crop-stage thresholds [RESEARCHED]. Relies on interpolated point/block data without deep terrain-aware physics downscaling [INFERENCE]. Agromet advisory rule structure inspiration [INFERENCE]. Downscaling model relies heavily on nearest-station interpolation rather than physical DEM/NDVI super-resolution [INFERENCE]. IMD Agrimet [RESEARCHED] [RESEARCHED]
ECMWF ERA5-Land / High-Res NWP (WRF) Physical atmospheric regional model running at ~9 km down to 3 km resolution [RESEARCHED]. Highly accurate physics-based fluid dynamics equations [KNOWN]. Computationally massive; requires hours on HPC clusters for single run; unviable for real-time hackathon API [INFERENCE]. Standard reference dataset for offline training [INFERENCE]. Cannot run dynamically on demand at scale for 2,60,000 Panchayats without massive cloud expenditure [INFERENCE]. ECMWF Docs [RESEARCHED] [RESEARCHED]
Private Commercial Apps (Skymet / Cropin) Commercial agri-tech platforms offering hyper-local forecasts and farm management [KNOWN]. Sleek UI, mobile-first design, push notifications [KNOWN]. Proprietary closed-source black-box models; expensive subscription fees; not integrated with official IMD/ICAR advisory channels [INFERENCE]. UX inspiration [INFERENCE]. Non-standardized advisories, lack of official IMD governance alignment, high cost for smallholders [INFERENCE]. Public Industry Reports [KNOWN] [KNOWN]

9. Candidate Solution Approaches

Approach 1: Geostatistical Downscaling with Lapse Rate & Kriging (GWR / EDK)

  • Mechanism: Applies physical temperature lapse rates (~6.5°C/km altitude increase) and Geographically Weighted Regression (GWR) / Kriging with External Drift (KED) using DEM elevation and distance-to-coast as covariates [RESEARCHED].
  • Problem Fit: High for temperature; moderate/low for convective monsoon rainfall [INFERENCE].
  • Accuracy Potential: Medium (RMSE ~1.2°C for Temp, higher error for intense localized rainfall) [INFERENCE].
  • Data Requirements: Low (Coarse forecast + SRTM DEM) [INFERENCE].
  • Complexity: Low (Deterministic statistical equations) [INFERENCE].
  • Explainability: High (100% physically interpretable formulas) [INFERENCE].
  • Cost: Very Low [INFERENCE].
  • Scalability: High (Fast computation) [INFERENCE].
  • Hackathon Feasibility: High [INFERENCE].
  • Production Feasibility: High (Low compute overhead) [INFERENCE].

Approach 2: Deep Learning Super-Resolution Architecture (Physics-Informed U-Net / SRCNN)

  • Mechanism: Treats spatial weather downscaling as an image super-resolution task. Low-resolution weather forecast grids (e.g., $16 \times 16$) are stacked with high-resolution static spatial channels ($128 \times 128$ DEM, Slope, Aspect, LULC, NDVI) and processed through a Physics-Informed U-Net / SRCNN architecture with conservation of mass/energy constraints [RESEARCHED].
  • Problem Fit: Very High (Captures non-linear spatial relationships and complex micro-terrain interactions) [INFERENCE].
  • Accuracy Potential: High ($R^2 > 0.92$, low RMSE across complex terrain) [RESEARCHED].
  • Data Requirements: High (Multi-year historical gridded weather + high-res satellite covariates) [INFERENCE].
  • Complexity: High [INFERENCE].
  • Explainability: Medium (Requires SHAP / Feature importance maps) [INFERENCE].
  • Cost: Low inference cost (GPU/CPU inference in milliseconds once trained) [INFERENCE].
  • Scalability: Very High [INFERENCE].
  • Hackathon Feasibility: Moderate-High (Pre-trained PyTorch model exported to ONNX) [INFERENCE].
  • Production Feasibility: High [INFERENCE].

Approach 3: Two-Stage Hybrid ML Pipeline (XGBoost Spatial Downscaler + Physics Lapse-Rate Post-Processing)

  • Mechanism: Stage 1 uses XGBoost regressor trained per grid-cell on terrain features (Elevation, Slope, Aspect, Distance to Water, NDVI, Centroid Lat/Lon) to predict local deltas ($\Delta Temp, \Delta Rain$) from block averages. Stage 2 applies physical mass-balance residual correction so the spatial integral of Panchayat forecasts matches the block-level forecast total [INFERENCE].
  • Problem Fit: High [INFERENCE].
  • Accuracy Potential: High ($R^2 \approx 0.90$) [INFERENCE].
  • Data Requirements: Moderate [INFERENCE].
  • Complexity: Moderate [INFERENCE].
  • Explainability: Very High (Feature importance + linear physics residual adjustment) [INFERENCE].
  • Cost: Very Low [INFERENCE].
  • Scalability: High [INFERENCE].
  • Hackathon Feasibility: Extremely High [INFERENCE].
  • Production Feasibility: High [INFERENCE].

10. Recommended Approach & Why

Selected Approach: Hybrid Physics-Informed Deep Super-Resolution Framework (Physics-Informed U-Net + Mass-Conservation Residual Adjustment)

[Low-Res Block Forecast (10-25km)] + [High-Res Spatial Covariates (30m DEM, Slope, Aspect, LULC, NDVI)]
                                      │
                                      ▼
             [Stage 1: Deep Super-Resolution Neural Network (U-Net)]
                                      │
                                      ▼
             [Stage 2: Physics-Guided Mass & Thermal Conservation Post-Processor]
                                      │
                                      ▼
           [High-Res Panchayat Weather Grid (1-3km) & LGD Vector Overlay]
                                      │
                                      ▼
             [Stage 3: Agromet Advisory Rules Engine (ICAR Matrix)]

Why Alternatives Were Rejected

  1. Pure Geostatistical Kriging (Approach 1) Rejected: Fails to capture non-linear convective rainfall dynamics during the Indian Summer Monsoon, treating micro-climates as purely linear elevation gradients [INFERENCE].
  2. Pure XGBoost Point Regression (Approach 3) Rejected: Treats each Panchayat cell independently without spatial context (convolutional receptive fields), missing mesoscale atmospheric movement and boundary continuity across adjacent Panchayats [INFERENCE].
  3. Pure Deep Learning Black-Box Model Rejected: Unconstrained deep neural networks can violate basic meteorological conservation laws (e.g., generating total downscaled Panchayat rainfall that exceeds or drops below total block atmospheric water content) [RESEARCHED].

Winning Synthesis: The recommended hybrid model uses U-Net for non-linear spatial feature extraction combined with a Physics-Guided Post-Processing Layer that enforces thermal lapse rate limits and conserves mass/water volume across the block boundary [INFERENCE].

11. Technical Architecture

Layered System Architecture Diagram

===================================================================================
1. INGESTION LAYER
   ├── IMD / GFS Block Forecast API Puller (Temperature, Rain, RH, Wind) [GRIB2/JSON]
   ├── Satellite & Spatial Covariate Ingestion (USGS 30m SRTM DEM, Sentinel-2 NDVI)
   └── LGD Code Gram Panchayat Vector Boundary Parser (GeoJSON/PostGIS)
===================================================================================
                                      │
                                      ▼
===================================================================================
2. PROCESSING & DOWN-SCALING PIPELINE (Python / PyTorch / ONNX)
   ├── Feature Raster Stacker (Coarse Weather Channel + Dense DEM/NDVI Channels)
   ├── Physics-Informed U-Net Super-Resolution Inference Engine (Block ➔ 1km Grid)
   └── Mass & Energy Balance Corrector (Ensures Block-Integral Consistency)
===================================================================================
                                      │
                                      ▼
===================================================================================
3. SPATIAL ANALYSIS & ADVISORY ENGINE
   ├── PostGIS Spatial Zonal Statistics (Raster-to-Panchayat Polygon Aggregation)
   ├── Agromet Rule Engine (Matching Downscaled Parameters vs ICAR Crop Thresholds)
   └── Advisory Content Generator (Generates Impact-based Advice + Vernacular Text)
===================================================================================
                                      │
                                      ▼
===================================================================================
4. STORAGE & CACHING LAYER
   ├── PostGIS Database (Panchayat Boundaries, Downscaled Forecast History, LGD Codes)
   ├── Redis In-Memory Cache (Pre-computed 5-day Panchayat Bulletins & Weather Grids)
   └── Object Storage / S3 (GeoTIFF Downscaled Maps & Crop Advisory PDF Bulletins)
===================================================================================
                                      │
                                      ▼
===================================================================================
5. API & DELIVERY LAYER (FastAPI / Node.js)
   ├── REST & GraphQL API Endpoints for Mobile/Web Client
   ├── Webhook & Push Trigger Engine (WhatsApp API / Firebase FCM / mKisan SMS)
   └── GeoServer / MapLibre Vector Tile Server (Panchayat Heatmap Visualization)
===================================================================================
                                      │
                                      ▼
===================================================================================
6. PRESENTATION LAYER (Client Interfaces)
   ├── Web Dashboard for DAMU / AMFU Scientists (Forecast Override & Verification)
   └── Offline-First Mobile PWA for Farmers (Voice Alerts, Advisory, Map View)
===================================================================================

Data Flow Narrative (Step-by-Step)

  1. Trigger & Fetch: Every 12 hours (or on IMD bulletin release at 11 AM/4 PM), the Ingestion Layer triggers a cron task fetching the latest 5-day Block-level forecast JSON/GRIB2 files [INFERENCE].
  2. Preprocessing & Stacking: The backend normalizes weather parameters and aligns them with pre-processed static raster layers (30m SRTM DEM elevation, slope, aspect, ESA WorldCover, MODIS NDVI) bounded by the target block [INFERENCE].
  3. Model Inference: The tensor tensor array ($C_{coarse} + C_{static}$) is passed into the pre-trained Physics-Informed U-Net model, outputting a high-resolution 1 km spatial grid for Max Temp, Min Temp, Rainfall, Relative Humidity, and Wind Speed [INFERENCE].
  4. Physics Correction: A post-processing kernel computes the spatial mean over the block; if the downscaled mean deviates from the original block forecast by $> \epsilon$, an isotropic residual bias matrix is applied to preserve mass/energy conservation [INFERENCE].
  5. Polygon Zonal Statistics: PostGIS executes spatial zonal averaging over each Gram Panchayat vector polygon using its unique LGD code, producing discrete 5-day weather numerical series per Panchayat [INFERENCE].
  6. Agromet Rule Evaluation: The downscaled Panchayat parameters are evaluated against the crop-stage rule matrix (e.g., If Crop = Paddy, Growth Stage = Flowering, Downscaled Rain > 25mm ➔ Trigger Blast Disease Warning & Postpone Fertilizer Application) [INFERENCE].
  7. Distribution: The final downscaled weather and advisory bulletin is written to Redis/PostgreSQL and pushed via REST/Webhooks to the farmer mobile app, WhatsApp bot, and DAMU scientific review portal [INFERENCE].

12. Technology Stack

Layer Selected Technology Why It Fits Alternative Considered Why Alternative Rejected
Frontend UI React / Next.js (TailwindCSS + MapLibre GL) PWA support, fast SSR rendering, excellent MapLibre vector tile support for high-res map overlays [INFERENCE]. Flutter / React Native Higher setup complexity for web-based judge demo; PWA works across desktop and mobile without app store delay [INFERENCE].
Backend API FastAPI (Python 3.12) Native async performance, direct integration with NumPy, PyTorch, GeoPandas, and Rasterio libraries [INFERENCE]. Node.js (Express) Poor native geospatial and ML tensor processing support; required wrapping Python scripts [INFERENCE].
Geospatial Engine PostGIS + Rasterio / GDAL Industry standard for vector/raster zonal statistics, spatial joins (ST_SummaryStats), and LGD boundary querying [RESEARCHED]. QGIS Server / ArcGIS API Heavy system overhead, licensing costs, non-programmatic for real-time API execution [INFERENCE].
AI / Downscaling ML PyTorch 2.x + ONNX Runtime PyTorch allows custom physics-loss functions during training; ONNX Runtime accelerates inference 4x on standard CPU servers [RESEARCHED]. TensorFlow 2.x Slower execution, heavier dependencies, less flexible custom spatial loss implementation [INFERENCE].
Database PostgreSQL 16 + PostGIS extension Spatial database capable of storing vector boundary shapefiles, spatial indexes (GIST), and structured relational advisories [KNOWN]. MongoDB / Spatial Lack of robust native raster-vector spatial intersection math compared to PostGIS [INFERENCE].
In-Memory Cache Redis Sub-millisecond latency for retrieving pre-computed Panchayat 5-day forecast payloads [KNOWN]. Memcached Lacks data structures like Geospatial Hashes and JSON store [INFERENCE].
Offline Sync / Storage Workbox Service Worker + IndexedDB (Dexie.js) Enables offline-first PWA operation in zero-connectivity rural Panchayats [INFERENCE]. LocalStorage LocalStorage is synchronous, blocking, and limited to 5MB storage limit [INFERENCE].
Notifications & Messaging Twilio / Gupshup WhatsApp API + Firebase FCM High delivery rate for vernacular text/audio advisories directly to rural mobile phones [RESEARCHED]. Pure SMS (mKisan) SMS length constraints (160 chars) prevent rich crop-stage advice and multi-image alerts [INFERENCE].

13. AI/ML Justification

Model 1: Physics-Informed Super-Resolution Downscaler (U-Net Architecture)

  1. Exact Problem Solved: Disaggregates coarse $12 \text{ km} \times 12 \text{ km}$ block forecast grids into $1 \text{ km} \times 1 \text{ km}$ spatial weather grids [INFERENCE].
  2. Input: Stacked multi-channel spatial tensor: Coarse block weather grid + High-res DEM + Slope + Aspect + LULC + NDVI ($16 \times 16 \times C \rightarrow 128 \times 128 \times C$) [INFERENCE].
  3. Output: High-resolution spatial map ($128 \times 128$) for 2m Temperature, Precipitation, Relative Humidity, Wind Speed [INFERENCE].
  4. Why Simpler Rule Cannot Solve: Simple bilinear/bicubic interpolation produces smooth gradients that ignore mountain ridges, valley temperature inversions, and micro-terrain rainfall barriers [RESEARCHED].
  5. Data Required: 5+ years of ERA5-Land (9 km) reanalysis, IMD 0.25° gridded daily rainfall/temp, and 30m SRTM DEM [RESEARCHED].
  6. Evaluation: Root Mean Squared Error (RMSE), Structural Similarity Index (SSIM), Spatial Correlation Coefficient ($R^2$), and Mass Conservation Error [RESEARCHED].
  7. What Happens When Wrong: Physics-guided post-processing safety layer overrides unrealistic extremes (e.g., temperature dropping below freezing in tropical plains) [INFERENCE].
  8. Requires Training: Yes (Offline supervised training on historical grid pairs) [INFERENCE].
  9. Requires Fine-Tuning: Yes (Per agro-climatic region if terrain complexity varies significantly) [INFERENCE].
  10. Requires RAG: No.
  11. Requires Computer Vision: Yes (Convolutional spatial operations on raster image bands) [INFERENCE].
  12. Requires Speech Processing: No.
  13. Requires Forecasting: No (It performs spatial downscaling on an already generated temporal forecast) [INFERENCE].
  14. Human Oversight: Yes; AMFU scientists review flag warnings before official bulletin release [INFERENCE].

Model 2: Multilingual Vernacular Voice Advisory Generator (Whisper + Indic-TTS)

  1. Exact Problem Solved: Converts structured text Agromet Advisories into localized regional voice clips for low-literacy farmers [INFERENCE].
  2. Input: Structured text advisory JSON (Hindi, Marathi, Tamil, Telugu, etc.) [INFERENCE].
  3. Output: Compressed MP3 audio file [INFERENCE].
  4. Why Simpler System Cannot Solve: Standard web browsers lack natural-sounding Indian accent regional TTS engines [INFERENCE].
  5. Data Required: Pre-trained AI4Bharat Indic-TTS / Bhashini API models [RESEARCHED].
  6. Evaluation: Mean Opinion Score (MOS) for speech clarity [INFERENCE].
  7. What Happens When Wrong: Visual text and icon fallback provided alongside audio [INFERENCE].
  8. Requires Training: No (Uses pre-trained Bhashini / AI4Bharat open models) [RESEARCHED].
  9. Requires Fine-Tuning: No.
  10. Requires RAG: No.
  11. Requires Computer Vision: No.
  12. Requires Speech Processing: Yes (Text-To-Speech) [INFERENCE].
  13. Requires Forecasting: No.
  14. Human Oversight: Optional; static template verification [INFERENCE].

Explicit Non-AI Boundaries

  • DO NOT use AI for Zonal Vector Aggregation: Polygon raster extraction MUST use deterministic exact spatial mathematics (PostGIS ST_SummaryStats / GeoPandas) [INFERENCE].
  • DO NOT use LLMs for Agromet Risk Thresholds: Crop stress decisions (e.g., "paddy pest risk threshold = 85% RH for 3 days") MUST be hard-coded from scientific ICAR documentation in deterministic SQL/JSON rule tables to prevent hallucinated agronomic advice [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Database Schema (Entity-Relationship Design)

-- 1. Gram Panchayat Administrative Units (LGD Code mapped)
CREATE TABLE gram_panchayats (
    lgd_code INT PRIMARY KEY,
    panchayat_name VARCHAR(150) NOT NULL,
    block_id INT NOT NULL,
    district_id INT NOT NULL,
    state_id INT NOT NULL,
    geom GEOMETRY(MultiPolygon, 4326), -- WGS84 Spatial Boundary
    centroid GEOMETRY(Point, 4326),
    elevation_mean_m FLOAT,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);
CREATE INDEX idx_gp_geom ON gram_panchayats USING GIST(geom);
CREATE INDEX idx_gp_lgd ON gram_panchayats(lgd_code);

-- 2. Downscaled Weather Forecast Data
CREATE TABLE downscaled_panchayat_forecasts (
    id BIGSERIAL PRIMARY KEY,
    lgd_code INT REFERENCES gram_panchayats(lgd_code),
    forecast_run_time TIMESTAMP NOT NULL,
    forecast_date DATE NOT NULL, -- 1-to-5 day outlook
    temp_max_c FLOAT NOT NULL,
    temp_min_c FLOAT NOT NULL,
    rainfall_mm FLOAT NOT NULL,
    rh_percent FLOAT NOT NULL,
    wind_speed_kmh FLOAT NOT NULL,
    model_version VARCHAR(50) DEFAULT 'PI-UNet-v1.2',
    confidence_score FLOAT DEFAULT 0.92,
    UNIQUE(lgd_code, forecast_run_time, forecast_date)
);
CREATE INDEX idx_fc_gp_date ON downscaled_panchayat_forecasts(lgd_code, forecast_date);

-- 3. Crop-Weather Advisory Rules Matrix
CREATE TABLE agromet_rules (
    rule_id SERIAL PRIMARY KEY,
    crop_name VARCHAR(100) NOT NULL,
    growth_stage VARCHAR(100) NOT NULL,
    parameter_name VARCHAR(50) NOT NULL, -- e.g., 'rainfall_mm', 'temp_max_c'
    operator VARCHAR(10) NOT NULL,        -- '>', '<', 'BETWEEN'
    threshold_value_1 FLOAT NOT NULL,
    threshold_value_2 FLOAT,
    advisory_impact TEXT NOT NULL,
    recommended_action TEXT NOT NULL,
    severity_level VARCHAR(20) -- 'INFO', 'WARNING', 'CRITICAL'
);

-- 4. Generated Panchayat Agromet Bulletins
CREATE TABLE panchayat_advisory_bulletins (
    bulletin_id BIGSERIAL PRIMARY KEY,
    lgd_code INT REFERENCES gram_panchayats(lgd_code),
    issue_date DATE NOT NULL,
    crop_name VARCHAR(100) NOT NULL,
    advisory_text_en TEXT NOT NULL,
    advisory_text_vernacular JSONB, -- {"hi": "...", "mr": "..."}
    audio_url VARCHAR(255),
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

Offline Caching, Bandwidth & Sync Architecture

  1. Client Caching Layer: Mobile PWA uses Dexie.js (IndexedDB) to store downscaled forecasts and advisories for the farmer's registered Panchayat code [INFERENCE].
  2. Bandwidth Optimization: Weather forecast REST payloads are minified JSON containing compressed 5-day arrays ([temp_max, temp_min, rain, rh, wind]), reducing payload size to < 2.5 KB per Panchayat [INFERENCE].
  3. Map Tile Optimization: Map overlays use Vector Tiles (.pbf format) generated by PostGIS/ST_AsMVT, compressing spatial boundary representations by 90% compared to raw GeoJSON [INFERENCE].
  4. Background Sync: Service Workers automatically queue user profile updates or farmer feedback while offline, executing HTTP POST synchronization when network connection is re-established [INFERENCE].

15. Security, Privacy & Governance

  • Personal & Location Data Protection: Farmer registration stores only the Gram Panchayat LGD code and preferred crop type, strictly avoiding precise GPS coordinate tracking to ensure zero location privacy risk [INFERENCE].
  • Authentication & Role-Based Access Control (RBAC):
    • Public / Farmers: Unauthenticated read-only access to downscaled Panchayat weather and advisory feeds [INFERENCE].
    • AMFU / DAMU Scientists: OAuth2 / JWT authentication with role-based write access to modify, approve, or issue advisory bulletins [INFERENCE].
  • Data Encryption: TLS 1.3 in transit; AES-256 for database storage [KNOWN].
  • Ownership & Open Government Data Policy: Raw weather downscaled rasters remain property of MoES / IMD. Output APIs released under Open Government Data (OGD) License India [RESEARCHED].
  • Liability & Prototype Disclaimer: AI downscaled forecasts are probabilistic estimations. Advisories issued via prototype MUST include an explicit governance disclaimer: "Downscaled advisory produced by AI prototype under SIH 2026. Official guidance provided by local Krishi Vigyan Kendra (KVK)" [INFERENCE].

16. MVP Scope (Hackathon Build)

+-----------------------------------------------------------------------+
|                         HACKATHON MVP SCOPE                           |
+-----------------------------------------------------------------------+
| [MUST BUILD]                                                          |
| ├── Downscaling AI pipeline for 1 Pilot State (e.g., Maharashtra)     |
| ├── Pre-trained Physics-Informed U-Net / XGBoost Downscaler           |
| ├── LGD Panchayat Vector Boundary Integration (PostGIS)               |
| ├── Automated Zonal Extraction for 500+ Gram Panchayats               |
| ├── ICAR Rule Engine generating crop advisories for 3 major crops     |
| └── Interactive Web Map Dashboard (MapLibre vector heatmap)           |
|                                                                       |
| [SHOULD BUILD]                                                        |
| ├── Multilingual Voice Advisory synthesis (Bhashini API integration)  |
| ├── Offline PWA mobile interface with IndexedDB caching               |
| └── WhatsApp Bot simulated alert trigger                              |
|                                                                       |
| [NICE TO HAVE]                                                        |
| ├── AMFU Scientist Manual Forecast Correction/Override Interface     |
| └── Historical forecast verification scorecards (RMSE / SSIM graphs)  |
+-----------------------------------------------------------------------+

Explicitly Out of Scope for Hackathon Build

  • Live physical AWS/ARG hardware deployment across 2,60,000 Panchayats [INFERENCE].
  • Real-time supercomputer integration (Arunika HPC cluster API connection) [INFERENCE].
  • Pan-India real-time inference for all 28 States simultaneously (limited to 1-2 pilot states during 36-hour build window) [ASSUMPTION].

17. Live Demo Simulation Plan

To present a convincing live demo during SIH judging without relying on closed government APIs or unavailable live satellite feeds, the team will implement an authentic, real-data-backed simulation pipeline:

[Simulated Real-Time Pipeline for SIH Demo]
  ├── Step 1: Pre-cached Real GFS/ERA5-Land 0.25° Weather Grids (Maharashtra Pilot)
  ├── Step 2: Live UI Trigger ➔ User selects any Gram Panchayat (e.g., Shirur Block)
  ├── Step 3: Real-Time Execution of PyTorch Downscaling Inference (1.2 seconds)
  ├── Step 4: Live Spatial Overlay of 1km Downscaled Grid on MapLibre Vector Map
  ├── Step 5: PostGIS Zonal Stat Extraction ➔ Displays Panchayat Weather Parameters
  └── Step 6: ICAR Rules Execution ➔ Renders Multilingual Text + Vernacular Voice Alert
  • Simulated Component 1 (Live Block Forecast Feed): High-resolution pre-downloaded ERA5-Land / GFS GRIB2 forecast files stored locally, loaded through a mock IMD REST API endpoint preserving real payload structures [ASSUMPTION].
  • Simulated Component 2 (WhatsApp Dissemination): Integrated Sandbox WhatsApp API (Twilio Sandbox) sending real-time advisory messages to judges' smartphones live during demo [ASSUMPTION].
  • Transparency Commitment: The team will explicitly state: "Downscaling AI model and PostGIS zonal extractions run LIVE during this demo. Input block forecasts are pulled from cached IMD/GFS open datasets." [INFERENCE]

18. Implementation Roadmap

Target Deadline: 30 September 2026 [PS]

[Phase 1: Research & Data Prep]  --->  [Phase 2: ML Model Training]  --->  [Phase 3: Backend & PostGIS]  --->  [Phase 4: UI & Advisory Engine]  --->  [Phase 5: Testing & SIH Demo]
      (Weeks 1 - 2)                           (Weeks 3 - 4)                       (Weeks 5 - 6)                       (Weeks 7 - 8)                      (Final Review)

Phase Details

  • Phase 1: Data Acquisition & Preprocessing (Target: 15 July 2026)

    • Download USGS 30m SRTM DEM, ESA WorldCover, and MODIS NDVI rasters for pilot state [KNOWN].
    • Download LGD Gram Panchayat GeoJSON boundaries [RESEARCHED].
    • Build dataset of paired coarse (ERA5 9km) vs fine (ERA5-Land / High-res AWS) historical weather grids [RESEARCHED].

  • Phase 2: AI Downscaling Model Architecture & Training (Target: 05 August 2026)

    • Implement PyTorch U-Net with custom Mass-Conservation Loss function [INFERENCE].
    • Train and validate model against historical test set; achieve $R^2 > 0.88$ for Temp and Rain [INFERENCE].
    • Export trained model weights to ONNX format for zero-GPU lightweight server inference [INFERENCE].

  • Phase 3: Spatial Engine & Backend Integration (Target: 25 August 2026)

    • Deploy PostgreSQL/PostGIS database and populate LGD vector boundary tables [INFERENCE].
    • Develop FastAPI backend pipelines for ONNX downscaling and PostGIS zonal statistics [INFERENCE].
    • Integrate ICAR crop-weather rule table for automated advisory generation [RESEARCHED].

  • Phase 4: Frontend UI, Vernacular Voice & Mobile PWA (Target: 15 September 2026)

    • Build React + MapLibre interactive web dashboard for AMFU scientists [INFERENCE].
    • Implement farmer mobile PWA with offline caching and Bhashini voice output [INFERENCE].
    • Connect Twilio WhatsApp sandbox trigger [ASSUMPTION].

  • Phase 5: System Verification, Stress Testing & SIH Presentation Prep (Target: 25 September 2026)

    • Execute offline performance tests, false-positive advisory checks, and end-to-end user journey validation [INFERENCE].
    • Finalize live demo scripts, backup video recordings, and SIH slide deck before 30 September 2026 deadline [PS].

19. Testing & Validation Plan

Test Category Target Metrics / Test Setup Success Criteria
Functional Testing Input 10 different block forecasts across varied topography (plains, hills, coastal) [INFERENCE]. 100% successful generation of downscaled Panchayat weather and crop advisories without API failure [INFERENCE].
AI Model Accuracy Compare downscaled 1 km predictions against un-seen ground-truth AWS/ARG station observations [RESEARCHED]. Temperature $RMSE < 1.1^\circ\text{C}$; Rainfall Spatial Correlation $R^2 > 0.85$; Mass Conservation Error $< 2%$ [INFERENCE].
Performance Testing Concurrent API request load test using Locust (100 parallel user requests for downscaled bulletins) [INFERENCE]. API Response time $< 800\text{ ms}$ per Panchayat request; Redis cache hit response time $< 50\text{ ms}$ [INFERENCE].
False-Positive Advisory Test Inject abnormal synthetic weather spikes (e.g., 60°C temperature anomaly) [INFERENCE]. System physics-check filter catches anomaly, prevents false panic advisory, and flags alert for AMFU review [INFERENCE].
Offline PWA Test Disconnect network connection on client browser; request saved Panchayat advisory [INFERENCE]. PWA seamlessly loads cached forecast and advisory from IndexedDB with "Offline Mode" banner [INFERENCE].
Usability & Field Test Test UI with 5 non-technical users in Hindi/vernacular text and voice [INFERENCE]. Task completion rate $> 90%$ for checking crop-stage weather advice without developer assistance [INFERENCE].

20. Risk Register

Risk Likelihood Impact Severity Mitigation
High Model Error in Extremely Complex Mountainous Terrain High High High Incorporate elevation aspect gradients and local lapse-rate physical constraints into neural loss function [INFERENCE].
Unavailable / Rate-Limited Government IMD Live APIs High Medium High Implement automated fallback parser to Open-Meteo / GFS open global weather forecast feeds [INFERENCE].
LGD Gram Panchayat Vector Boundary Discrepancies Medium Medium Medium Pre-clean and topological-repair shapefiles using GDAL/OGR before importing into PostGIS [INFERENCE].
Hallucinated / Incorrect Agronomic Advice Causing Crop Loss Low High High Deterministic hard-coded ICAR rule table; strictly prohibit LLM dynamic text generation for advisory actions [INFERENCE].
High Latency in Real-Time Deep Learning Raster Inference Medium Medium Medium Export PyTorch model to ONNX Runtime; pre-calculate and cache 5-day downscaled rasters in Redis every 12 hours [INFERENCE].

21. Cost & Scalability

Hackathon Phase Costs (INR / USD)

  • Compute / Hosting: INR 0 / $0 (Utilizing local laptop GPU for model training, free-tier Vercel/Render for frontend/backend, free Cloud Supabase PostgreSQL/PostGIS) [ASSUMPTION].
  • Data Access: INR 0 / $0 (Free open datasets: USGS SRTM DEM, ERA5-Land, Open-Meteo, LGD boundary shapefiles) [KNOWN].
  • Messaging APIs: INR 0 / $0 (Twilio WhatsApp free developer sandbox) [ASSUMPTION].
  • Total Hackathon Cost: ₹0 (100% Free Open-Source Architecture) [INFERENCE].

Production Phase Deployment Costs (Pan-India Scale)

  • Cloud Infrastructure (AWS / GCP):
    • Database Server: AWS RDS PostgreSQL (db.r6g.xlarge with PostGIS) – ~$350/month [INFERENCE].
    • Inference Compute: 2x AWS EC2 c6i.2xlarge (8 vCPU, 16GB RAM) running ONNX Runtime – ~$280/month [INFERENCE].
    • Cache & Storage: AWS ElastiCache Redis + S3 Object Storage – ~$120/month [INFERENCE].
  • WhatsApp / Bulk SMS Push Dissemination: Gupshup / Kaleyra bulk SMS (100 million alerts/month at ₹0.12/SMS) – ~$140,000/month [INFERENCE].
  • Production Scalability: System scales horizontally by pre-computing downscaled rasters batch-wise every 12 hours. End-user reads are 100% served from Redis/CDN cache, allowing static infrastructure to serve 100+ million farmers at minimal server compute cost [INFERENCE].

22. Differentiation

Parameter Generic SIH Submissions Our Proposed Solution Requirements Justification
Downscaling Method Basic mathematical bilinear interpolation or random regression [INFERENCE]. Physics-Informed Super-Resolution Neural Network (U-Net) + Mass-Conservation Correction Layer [INFERENCE]. Accurately captures non-linear mountain/coastal micro-climates without violating atmospheric mass conservation [R1, R2].
Spatial Reference Generic GPS point estimates or arbitrary bounding boxes [INFERENCE]. Official LGD (Local Government Directory) Gram Panchayat Vector Boundaries [RESEARCHED]. Directly aligns downscaled weather outputs with official Indian administrative governance units [R1].
Agromet Intelligence Raw weather parameter charts without context [INFERENCE]. Stage-Specific ICAR Rule Decision Engine + Vernacular Audio Advisory [RESEARCHED]. Converts raw weather numbers into actionable, last-mile decision support for rural farmers [R3].
System Performance Heavy unoptimized Python scripts running slow raster math per user request [INFERENCE]. ONNX Runtime + PostGIS Zonal Stat Caching + Offline-First PWA [INFERENCE]. Sub-second response times with full offline accessibility in low-network Gram Panchayats [R1, R3].

23. Future Scope

  1. Hyper-Local Radar & Drone Integration: Incorporating real-time Doppler Weather Radar (DWR) reflectivity feeds and local IoT weather station streams for 1-hour nowcasting at Panchayat level [INFERENCE].
  2. Soil Moisture & Hydrological Coupling: Coupling downscaled precipitation outputs with the Noah-MP land surface model to predict soil moisture deficit and hyper-local crop water stress [RESEARCHED].
  3. Hyper-Targeted Crop Insurance Indexing: Providing verified downscaled weather logs to PMFBY (Pradhan Mantri Fasal Bima Yojana) for faster, claim-settlement verification at the Gram Panchayat unit level [RESEARCHED].

24. Assumptions & Open Items Log

Consolidated Assumptions

  • [ASSUMPTION] Historical ERA5-Land (9km) reanalysis coupled with IMD 0.25° grid datasets can act as a reliable surrogate for training the downscaling neural network during the hackathon.
  • [ASSUMPTION] The 36-hour hackathon MVP focus will be restricted to a single representative pilot state (e.g., Maharashtra or Uttar Pradesh) before pan-India expansion.
  • [ASSUMPTION] Twilio WhatsApp developer sandbox is sufficient for demonstrating last-mile alert triggers during judge evaluation.

Consolidated Open Items & Items Needing Validation

  • [NEEDS VALIDATION] Direct REST API credentials and rate limits for IMD's operational internal GKMS Block Forecast server.
  • [NEEDS VALIDATION] High-resolution topological boundary shapefiles for newly created Gram Panchayats updated in LGD during 2025–2026.

Final Recommendation

The team should PROCEED with building this solution. The combination of Physics-Informed Deep Super-Resolution (U-Net) and PostGIS LGD Zonal Aggregation directly solves the MoES / IMD problem statement (SIH26074) with high technical realism, actionable agricultural impact, and minimal cloud compute costs [INFERENCE].

Recommended Stack

  • AI/ML: PyTorch 2.x, ONNX Runtime, GeoPandas, Rasterio [INFERENCE].
  • Backend & DB: FastAPI (Python 3.12), PostgreSQL 16 + PostGIS, Redis [INFERENCE].
  • Frontend: React / Next.js, MapLibre GL JS, Workbox PWA, Dexie.js [INFERENCE].
  • Integrations: Bhashini Vernacular Speech API, Twilio WhatsApp Sandbox [RESEARCHED].

Biggest Dependency

Availability of static 30m SRTM DEM GeoTIFF rasters and clean GeoJSON shapefiles mapped to official Local Government Directory (LGD) Panchayat codes [INFERENCE].

Biggest Risk

Physics non-compliance in purely deep-learning downscaling outputs (e.g., generating unphysical localized extreme rainfall spikes) [RESEARCHED]. Mitigated via the proposed Mass-Conservation residual adjustment layer [INFERENCE].

Biggest Technical Challenge

Executing real-time spatial zonal statistics over thousands of Gram Panchayat polygons efficiently during user requests [INFERENCE]. Solved by pre-computing downscaled grids and caching aggregated Panchayat JSON payloads in Redis [INFERENCE].

First 5 Actions

  1. Download 30m SRTM DEM GeoTIFF and LGD Gram Panchayat GeoJSON boundary files for the selected pilot state (e.g., Maharashtra) [INFERENCE].
  2. Download 3 years of daily historical GFS/ERA5 coarse weather grids and fine ERA5-Land grids for offline ML dataset setup [RESEARCHED].
  3. Build and train baseline PyTorch U-Net downscaling model on local GPU; verify $R^2$ accuracy metrics [INFERENCE].
  4. Set up PostgreSQL / PostGIS spatial database and ingest LGD vector boundary tables [INFERENCE].
  5. Implement FastAPI backend route that takes low-res forecast input, runs ONNX downscaling model, and returns Panchayat JSON forecast [INFERENCE].

Validation Before Coding

  • Verify that the selected pilot state LGD shapefile loads cleanly into PostGIS without invalid polygon geometries (ST_IsValid) [INFERENCE].
  • Test that ONNX Runtime inference executes in $< 2.0\text{ seconds}$ on CPU for a single block domain [INFERENCE].

Buildability Score /100

92 / 100 (High buildability for a 4-6 student team using Python geospatial libraries, pre-trained UNet architectures, open SRTM DEM data, and standard PostGIS workflows) [INFERENCE].

Confidence

High [95%] (Supported by operational IMD downscaling research, public ERA5-Land/SRTM datasets, and standard geostatistical/deep-learning super-resolution methodologies) [RESEARCHED].

Research Sources

  1. Title: PIB Release: Parliament Question on Weather Forecasting Data Provided to Farmers by IMD

    • URL: https://www.pib.gov.in/PressReleasePage.aspx?PRID=2223075
    • Source Type: Official Government Document (PIB / MoES)
    • Supported Claim: IMD provides block-level weather forecasts and AAS twice weekly under GKMS scheme across 127 agroclimatic zones [1.1].
    • Reliability: Authoritative (Official MoES Statement)

  2. Title: PIB Release: AI-Enabled Systems Introduced by IMD for Hyper-Local Forecasts

  3. Title: KALP Platform Document - Agricultural Meteorology Division, IMD

    • URL: https://imdagrimet.gov.in/Files/Kalp_Sankalp/KALP.pdf
    • Source Type: Technical Manual / Government Report
    • Supported Claim: KALP platform enables location selection up to Gram Panchayat level and dynamically compares forecasts with crop growth stage thresholds [1.3].
    • Reliability: Authoritative

  4. Title: Research Article: Intercomparison of Machine Learning Models for Spatial Downscaling of Daily Mean Temperature in Complex Terrain

    • URL: https://www.mdpi.com/2073-4433/15/9/1085
    • Source Type: Peer-Reviewed Academic Publication (MDPI)
    • Supported Claim: CNN / U-Net convolutional architectures outperform traditional machine learning and statistical methods for spatial downscaling across complex terrain [4.10].
    • Reliability: High (Peer-Reviewed Scientific Literature)

  5. Title: Meghdoot Mobile Application - CGSpace / ICRISAT / IMD Documentation

02 / What our

DEVELOPERS SAY

" 30M+ Google Gemini tokens went into researching, analyzing, and understanding all 233 SIH problem statements to make them easier to explore and solve. "
Satyam Kumar
Satyam Kumar
CEO , Full Stack Developer
" The platform makes SIH problem statements much easier to understand. Having everything from the problem details to the procedure in one place. "
Rishabh Ranjan
Rishabh Ranjan
COO , Content Writer
" 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. "
Himanshu Kumar
Himanshu Kumar
CMO , Product Reviewer
03 / Are You

READY TO
BUILD?

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.

LinkedIn
[email protected]