AI-Driven Hyper-Local Early Warning System for Severe Weather Nowcasting

1. Executive Snapshot

This dossier provides a comprehensive technical, architectural, and strategic roadmap for developing an AI-Driven Hyper-Local Early Warning System for Severe Weather Nowcasting for the Ministry of Earth Sciences (MoES) and the National Centre for Medium Range Weather Forecasting (NCMRWF) [PS]. Traditional physics-based Numerical Weather Prediction (NWP) models suffer from high computational latency (often requiring hours to execute multi-node HPC simulations) and struggle to capture rapid, meso-beta and meso-gamma scale atmospheric fluctuations preceding severe localized weather [PS, KNOWN]. The proposed solution replaces computationally intensive thermodynamic simulations during the nowcasting window (2 to 6 hours) with a multi-modal spatiotemporal deep learning architecture (e.g., cross-attention multi-task spatiotemporal transformer) [PS, INFERENCE]. By ingesting real-time INSAT-3D/3DR geostationary satellite radiances (Water Vapor and Thermal Infrared channels), IMDAA high-resolution regional reanalysis thermodynamic baselines, and Digital Elevation Models (CartoDEM/SRTM), the engine simultaneously computes hyper-local probability risk maps for severe thunderstorms, cloudbursts, and flash floods [PS]. The platform combines a lightweight inference engine, an interactive GIS dashboard with Explainable AI (XAI) meteorological trigger heatmaps for disaster authorities, and an automated Common Alerting Protocol (CAP) compliant API for immediate first-responder dissemination [PS, INFERENCE].

Metadata Table

Fundamental System Objective:
To bypass NWP computational latency by deploying a multi-task spatiotemporal deep learning engine that ingests real-time INSAT-3D/3DR satellite imagery, IMDAA thermodynamic profiles, and high-resolution DEMs to simultaneously generate 2-to-6-hour hyper-local probability maps and automated alerts for severe thunderstorms, cloudbursts, and flash floods [PS].

2. Problem Deconstruction

The explicit problem statement text has been deconstructed into distinct atomic requirements. Every explicit clause, variable, and deliverable is retained without modification, weakening, or removal [PS].

Atomic Requirements Breakdown

1. [REQ-01] Computational Latency Bypassing:
   Source Text: "Traditional physics-based Numerical Weather Prediction (NWP) models often suffer from computational latency and struggle to capture the rapid, small-scale atmospheric changes... entirely bypassing the computational latency typical of traditional numerical weather prediction (NWP) models." [PS]
   Interpretation: The nowcasting system must deliver predictions within seconds-to-minutes of satellite data ingestion, avoiding multi-hour thermodynamic differential equation solvers [INFERENCE].
   Implementation Implication: Machine learning inference must run on GPU/accelerated infrastructure with optimized forward-pass latency (< 30 seconds per grid inference) [ASSUMPTION].
   Tag: [PS]

2. [REQ-02] 2 to 6 Hour Lead Time Nowcasting:
   Source Text: "...capable of 'nowcasting' severe weather 2 to 6 hours before impact, providing actionable lead time for disaster management." [PS]
   Interpretation: Output probabilities must forecast atmospheric state and risk intensity between $T+2\text{h}$ and $T+6\text{h}$ from the observation timestamp $T$ [INFERENCE].
   Implementation Implication: Spatiotemporal sequences (e.g., last 2 hours of 15-min satellite frames) must map to future multi-step prediction horizons up to $T+6\text{h}$ [INFERENCE].
   Tag: [PS]

3. [REQ-03] Simultaneous Multi-Hazard Nowcasting:
   Source Text: "...simultaneously predicts the onset of highly localized, rapidly intensifying events, namely severe thunderstorms, cloudbursts, and the subsequent flash floods..." [PS]
   Interpretation: The system must not run three isolated models; it must predict all three hazard classes in a unified pass to maintain physical consistency across linked phenomena [INFERENCE].
   Implementation Implication: Requires a Multi-Task Learning (MTL) architecture with a shared feature extraction backbone and distinct classification/regression heads [PS].
   Tag: [PS]

4. [REQ-04] Moisture Tracking via Integrated Water Vapor (IWV):
   Source Text: "Moisture Availability (The Fuel): The cornerstone of our storm nowcasting is the capture of integrated water vapor (IWV) variations. By tracking rapid spatial and temporal accumulations of IWV from satellites, the model identifies the concentrated moisture pools..." [PS]
   Interpretation: Spatial differential ($\nabla \text{IWV}$) and temporal rate of change ($\partial \text{IWV}/\partial t$) must be explicit input features or latent representations derived from INSAT-3D/3DR Water Vapor channels [INFERENCE].
   Implementation Implication: Satellite pre-processing pipelines must compute channel brightness temperatures and IWV proxies across consecutive satellite scans [INFERENCE].
   Tag: [PS]

5. [REQ-05] Instability Assessment (CAPE & CIN):
   Source Text: "Atmospheric Instability (The Energy): High Convective Available Potential Energy (CAPE) paired with eroding Convective Inhibition (CIN) serves as a prime indicator of impending severe thunderstorms." [PS]
   Interpretation: Thermodynamic profile variables from IMDAA reanalysis (temperature and humidity across pressure levels) must be ingested to derive or supply CAPE ($\text{J/kg}$) and CIN ($\text{J/kg}$) surfaces [INFERENCE].
   Implementation Implication: Data ingestion pipeline must extract multi-level atmospheric variables from IMDAA NetCDF/GRIB2 files and calculate/interpolate CAPE/CIN fields [INFERENCE].
   Tag: [PS]

6. [REQ-06] Kinematics, Lift, and Wind Shear Integration:
   Source Text: "Kinematics and Lift (The Trigger & Structure): Low-level convergence (wind vectors colliding at the surface) forces air upward... vertical wind shear (changes in wind speed/direction with altitude) helps the model predict whether a storm will move quickly or remain stationary." [PS]
   Interpretation: Surface wind components ($U_{10\text{m}}, V_{10\text{m}}$) for surface divergence/convergence ($\nabla \cdot \vec{V}$) and multi-level winds ($850\text{ hPa}, 500\text{ hPa}, 200\text{ hPa}$) for bulk shear vector calculation must be included [INFERENCE].
   Implementation Implication: Spatial derivative layers (Sobel/Laplacian or vector divergence kernels) must compute velocity convergence grids [INFERENCE].
   Tag: [PS]

7. [REQ-07] Cloud Top Temperature (CTT) Drop Rate:
   Source Text: "Observational Signatures: Rapid cooling of cloud tops, measured as the Cloud Top Temperature (CTT) Drop Rate, provides real-time validation of explosive vertical updrafts..." [PS]
   Interpretation: Continuous temporal change ($\Delta \text{CTT} / \Delta t$) derived from INSAT Thermal Infrared (TIR-1/TIR-2) brightness temperature differences across consecutive 15/30-minute intervals [INFERENCE].
   Implementation Implication: Sequence frames must be stored in a rolling buffer to calculate time-series derivatives for every pixel grid [INFERENCE].
   Tag: [PS]

8. [REQ-08] Topographic Overlay for Flash Flood Routing:
   Source Text: "Topographic Dynamics (The Flood Catalyst): To accurately predict flash floods, the AI overlays the atmospheric probability maps onto a high-resolution Digital Elevation Model (DEM)... calculate how terrain slope, elevation, and natural drainage basins will channel the extreme precipitation..." [PS]
   Interpretation: Atmospheric precipitation/cloudburst output must be coupled with surface hydrological routing (flow accumulation, slope, stream network) derived from CartoDEM/SRTM [INFERENCE].
   Implementation Implication: GIS hydrological processing module (e.g., TopoFlow / D8 flow direction algorithm) must convert predicted rainfall intensity into localized run-off inundation risk [INFERENCE].
   Tag: [PS]

9. [REQ-09] Explicit Data Source Ingestion:
   Source Text: Ingestion of (1) IMDAA Reanalysis Data (temperature, humidity profiles, geopotential height, U/V wind), (2) INSAT-3D/3DR Satellite Observations via MOSDAC (WV, TIR, QPE), and (3) Digital Elevation Model (CartoDEM or SRTM) [PS].
   Interpretation: System pipeline must ingest multi-modal raster and tabular data across differing spatial/temporal resolutions [INFERENCE].
   Implementation Implication: Spatiotemporal grid re-sampling, spatial alignment (CRS projection to WGS84 / EPSG:4326), and temporal interpolation pipeline required [INFERENCE].
   Tag: [PS]

10. [REQ-10] Multi-Task Spatiotemporal Transformer Architecture:
    Source Text: "...shared multi-modal spatiotemporal transformer network continuously analyzes real-time satellite grids... using cross-attention mechanisms. Utilizing a Multi-Task Learning (MTL) architecture, the network branches into distinct output 'heads.'" [PS]
    Interpretation: Model must use a Transformer backbone incorporating cross-attention between satellite imagery and thermodynamic grids, splitting into three distinct head branches [INFERENCE].
    Implementation Implication: Deep learning network implemented in PyTorch/TensorFlow using spatiotemporal attention blocks with multi-loss objective functions [INFERENCE].
    Tag: [PS]

11. [REQ-11] Interactive Web Spatial Dashboard with XAI:
    Source Text: "...interactive, webbased spatial dashboard designed for disaster management authorities, featuring dynamic risk maps overlaid on a Digital Elevation Model (DEM) and an Explainable AI (XAI) module that transparently displays meteorological triggers." [PS]
    Interpretation: Web UI must render vector/raster map layers and provide feature importance/gradient saliency maps (e.g., Integrated Gradients / SHAP maps showing IWV surge or CTT drop) [INFERENCE].
    Implementation Implication: React / Mapbox GL JS frontend with dynamic layer toggles and an XAI visual panel [INFERENCE].
    Tag: [PS]

12. [REQ-12] Automated Categorized Alerting API:
    Source Text: "...automated API will translate these predictive insights into immediate, categorized alerts sent directly to first responders and vulnerable communities..." [PS]
    Interpretation: Server backend must generate structured alert payloads (JSON/CAP format) and push notifications when threshold limits are exceeded [INFERENCE].
    Implementation Implication: FastAPI REST endpoint emitting Common Alerting Protocol (CAP v1.2) XML/JSON feeds and WebSocket/WebPush triggers [INFERENCE].
    Tag: [PS]

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Geographic & Orographic Realities:
   Mountainous vs. Plain Terrain: Cloudbursts occur primarily in steep Western Himalayan terrain (Himachal Pradesh, Uttarakhand, Jammu & Kashmir) and Western Ghats, where rapid orographic lift triggers explosive convection [KNOWN]. Elevation change across a single 10 km grid can exceed 2,000 meters, demanding high-resolution DEM modeling [KNOWN].

2. Satellite Temporal & Spatial Resolution Constraints:
   INSAT-3D/3DR Cadence: Geostationary images from INSAT-3D/3DR Imager are acquired every 30 minutes in standard mode or 15 minutes in interleaved mode [RESEARCHED]. Spatial resolution is 1 km for Visible, 4 km for Thermal IR (TIR), and 8 km for Water Vapor (WV) channels [RESEARCHED].
   Resolution Mismatch: Atmospheric predictors (WV channel at 8 km) must be fused with high-resolution DEM (CartoDEM at 30 m), requiring multi-scale spatial interpolation and grid alignment [INFERENCE].

3. Reanalysis vs. Operational Data Latency:
   IMDAA Reanalysis Limits: IMDAA is a historical reanalysis dataset (1979–2020) [RESEARCHED]. In an operational nowcasting environment, real-time thermodynamic baselines must be sourced from operational NWP analysis feeds (such as NCMRWF NCUM operational analysis or IMD GFS analysis at 12 km resolution) [RESEARCHED, INFERENCE].

4. Field Connectivity & Infrastructure Constraints:
   Last-Mile Connectivity: High-altitude Himalayan valleys frequently lose cellular broadband (4G/5G) and grid power during severe convective storms [KNOWN].
   Impact: Emergency alerts sent to first responders must rely on lightweight, highly compressed payloads (e.g., CAP SMS / Cell Broadcast) rather than bandwidth-heavy GIS video maps [INFERENCE].

5. Digital Literacy & User Interface Realities:
   Disaster Managers vs. Citizens: Field workers and village disaster committees require simple color-coded threat levels (Green, Yellow, Orange, Red) and local language text (Hindi/State languages) rather than raw meteorological metric charts [KNOWN].

6. Research Questions

1. Data Processing: How can 8 km INSAT Water Vapor brightness temperatures be spatially aligned and super-resolved with 12 km IMDAA thermodynamic grids and 30 m CartoDEM elevation rasters without introducing smoothing artifacts? [INFERENCE]
2. Satellite Latency: What is the actual end-to-end latency between INSAT-3D/3DR scan acquisition, MOSDAC data portal publication, and API access availability? [RESEARCHED]
3. Hydrological Coupling: What hydro-routing approximation (e.g., Height Above Nearest Drainage - HAND vs. D8 Flow Accumulation) offers sub-second computational speed during multi-task neural network inference? [RESEARCHED]
4. XAI Interpretability: Which feature attribution technique (Integrated Gradients, Grad-CAM, or SHAP) generates meteorologically coherent saliency maps for non-expert disaster authorities without slowing real-time execution? [INFERENCE]
5. Multi-Task Loss Balancing: How should multi-task loss terms (Binary Cross-Entropy for thunderstorm onset, Focal Loss for rare cloudburst events, and MSE/Dice Loss for flood extent) be dynamically weighted during model training? [INFERENCE]

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

+-----------------------------------------------------------------------------------+
|                        EXISTING WEATHER & DISASTER SYSTEMS IN INDIA               |
+-----------------------------------------------------------------------------------+
|  1. IMD Nowcast Portal (District/Station-based, manual forecaster rules)          |
|  2. South Asia Flash Flood Guidance System (SAsiaFFGS - WMO/IMD, watershed level)   |
|  3. Damini App (IITM Pune, lightning sensor network tracking)                     |
|  4. NDMA SACHET Platform (CAP alert aggregator & SMS dissemination hub)           |
+-----------------------------------------------------------------------------------+
                                          |
                                    REMAINING GAPS
                                          v
+-----------------------------------------------------------------------------------+
|  - High computational latency in physics-based NWP models (>3-6 hours)            |
|  - Spatial coarseness: Warnings issued at district level, not hyper-local (1-4km) |
|  - Siloed operations: Lightning, rain, and floods handled by disconnected apps    |
|  - Lack of XAI: Black-box outputs without visible atmospheric precursors          |
+-----------------------------------------------------------------------------------+

Comparative Analysis of Existing Systems

1. IMD Nowcast (District / Station Level)

   • Description: IMD issues 3-hourly nowcasts for districts using Doppler Weather Radar (DWR) and automatic weather stations [RESEARCHED].
   • Strengths: Operational, widely trusted, integrated with national disaster alerts [KNOWN].
   • Weaknesses: Issued at district administrative scales; misses micro-scale mountain cloudbursts; reliant on ground DWR range (<250 km) which has radar coverage gaps in high-altitude Himalayan valleys [RESEARCHED, KNOWN].
   • Relevance: Baseline standard for operational nowcasts in India [KNOWN].
   • Gap: Lack of sub-district (1–4 km) spatial precision and lack of integrated terrain hydrological routing for automated flash flood prediction [INFERENCE].
   • Confidence: [RESEARCHED]

2. South Asia Flash Flood Guidance System (SAsiaFFGS)

   • Description: WMO/IMD operational system providing flash flood threat guidance at the watershed scale using land surface models and NWP rain forecasts [RESEARCHED].
   • Strengths: Hydrologically sound watershed-level risk estimation [RESEARCHED].
   • Weaknesses: Relies heavily on global/regional NWP rain inputs (WRF/NCUM) which suffer from precipitation timing and location biases during extreme convection [RESEARCHED].
   • Relevance: Direct precursor for flash flood guidance [RESEARCHED].
   • Gap: Does not utilize real-time deep learning cross-attention on satellite Water Vapor (IWV) dynamic accumulation and CTT drop rates [PS, INFERENCE].
   • Confidence: [RESEARCHED]

3. Damini Lightning App (IITM Pune / MoES)

   • Description: Mobile app providing real-time lightning alerts based on ground-based Lightning Location Networks (LLN) [RESEARCHED].
   • Strengths: Highly accurate ground-truth detection of active cloud-to-ground lightning strikes [RESEARCHED].
   • Weaknesses: Diagnostic rather than predictive; alerts users after lightning strikes have begun; does not predict cloudbursts or flash floods [INFERENCE].
   • Relevance: Handles thunderstorm/lightning hazard aspect [RESEARCHED].
   • Gap: Single-hazard focus; no 2-to-6-hour predictive lead window before convective ignition [PS, INFERENCE].
   • Confidence: [RESEARCHED]

4. NDMA SACHET (Common Alerting Protocol Portal)

   • Description: Pan-India early warning aggregation portal developed by C-DOT and NDMA utilizing CAP v1.2 for LB-SMS and Cell Broadcast [RESEARCHED].
   • Strengths: Massive reach across telecom operators; multi-lingual alert delivery [RESEARCHED].
   • Weaknesses: An alerting engine, not a predictive meteorological modeling engine; depends entirely on upstream agencies to generate warnings [INFERENCE].
   • Relevance: Target target API endpoint for our proposed system's alerts [INFERENCE].
   • Gap: Our system fills the upstream predictive gap by feeding high-precision hyper-local risk polygons into SACHET via standard CAP APIs [PS, INFERENCE].
   • Confidence: [RESEARCHED]

9. Candidate Solution Approaches

To determine the optimal architecture for a 4–6 student team, three viable technical approaches were evaluated:

+--------------------------------------------------------------------------------------------------+
|                                    CANDIDATE APPROACHES EVALUATION                               |
+--------------------------------------------------------------------------------------------------+
|  Approach 1: Physics-Heuristic Rule Engine + GIS Overland Flow                                   |
|  - Threshold rules on CAPE/CIN & CTT; classical D8 flow routing                                  |
|  - Low compute, high explainability, BUT high false alarms & poor non-linear convective capture  |
+--------------------------------------------------------------------------------------------------+
|  Approach 2: Separate Classical ML / CNN Models per Hazard                                       |
|  - 3 independent models (XGBoost for Thunderstorm, ConvLSTM for Cloudburst, HAND for Flood)     |
|  - Modular build, BUT high pipeline complexity, redundancy, no cross-hazard feature sharing     |
+--------------------------------------------------------------------------------------------------+
|  Approach 3: Multi-Task Spatiotemporal Cross-Attention Transformer (RECOMMENDED)                |
|  - Unified encoder backbone with cross-attention across Satellite & IMDAA grids                  |
|  - 3 multi-task heads for simultaneous prediction + CartoDEM routing overlay                     |
|  - Best accuracy, direct match to PS requirements, low inference latency (<15s)                  |
+--------------------------------------------------------------------------------------------------+

Comparative Evaluation Matrix

10. Recommended Approach & Why

The recommended approach is Approach 3: Multi-Task Spatiotemporal Cross-Attention Transformer Engine with Hydrological Topographic Coupling [PS, INFERENCE].

Why Alternatives Were Rejected

1. Rejection of Approach 1 (Heuristic Rule Engine):
   Severe convective events (cloudbursts > 100 mm/hr) are non-linear atmospheric instabilities [KNOWN]. Fixed threshold rules (e.g., "CAPE > 2000 AND CTT drop > 10°C") yield unacceptably high false positive rates because high CAPE often persists without triggering lift mechanism, or lift occurs under severe CIN capping [KNOWN, INFERENCE]. It fails the core problem objective [PS].

2. Rejection of Approach 2 (Decoupled Separate Models):
   Running three separate machine learning pipelines (one for thunderstorms, one for cloudbursts, one for floods) creates redundant data ingestion loops, triples GPU memory overhead, and ignores the intrinsic physical feedback loops (e.g., cloudburst moisture flux directly drives flash flood hydrographs) [INFERENCE]. Furthermore, it violates the explicit PS requirement: "To predict multiple extreme events simultaneously, the engine employs a multi-task learning approach" [PS].

Why Approach 3 Wins

• Exact Problem Match: Faithfully mirrors every explicit element in SIH26077 (IWV tracking, CTT drop rate, CAPE/CIN instability, low-level convergence, multi-task learning heads, DEM flood channeling, XAI visual dashboard) [PS].
• Feature Synergy: The shared encoder backbone learns foundational atmospheric dynamics (moisture accumulation, thermal instability, wind vector convergence) that benefit all three hazard heads simultaneously, improving generalized accuracy over single-task models [INFERENCE].
• Inference Speed: Executing a single forward pass on an ONNX-optimized transformer takes < 15 seconds on a single GPU (e.g., NVIDIA T4 / RTX 4090), providing true real-time nowcasting with zero computational bottleneck [ASSUMPTION].

11. Technical Architecture

Layered System Architecture Diagram

+---------------------------------------------------------------------------------------------------+
|                                      SYSTEM ARCHITECTURE                                          |
+---------------------------------------------------------------------------------------------------+

[LAYER 1: DATA INGESTION & SATELLITE PIPELINE]
  +-----------------------+     +-----------------------+     +-----------------------+
  | INSAT-3D/3DR (MOSDAC) |     | IMDAA Reanalysis      |     | CartoDEM / SRTM       |
  | - WV (6.5-7.1 µm)     |     | - CAPE/CIN, Temp      |     | - 30m Elevation       |
  | - TIR1 (10.2-11.2 µm) |     | - U/V Winds, Geopot.  |     | - Slope & Drainage    |
  | - QPE Rain Estimates  |     | - Specific Humidity   |     | - D8 Flow Directions  |
  +-----------+-----------+     +-----------+-----------+     +-----------+-----------+
              |                             |                             |
              +-----------------------------+-----------------------------+
                                            v
[LAYER 2: SPATIOTEMPORAL DATA FUSION & ALIGNMENT GRID ENGINE]
  +-----------------------------------------------------------------------------------+
  | - Reprojection to EPSG:4326 (WGS84) & Spatial Interpolation to Unified 0.04° Grid |
  | - Continuous Temporal Rate Calculations: ΔIWV/Δt and ΔCTT/Δt                       |
  | - Divergence & Shear Calculators: (∂U/∂x + ∂V/∂y) and Bulk Wind Shear             |
  | - Tensor Normalization & Rolling 2-Hour Window Stacking [B, T=4, C=12, H, W]       |
  +-----------------------------------------+-----------------------------------------+
                                            v
[LAYER 3: AI/ML INFERENCE ENGINE (PyTorch / ONNX Runtime)]
  +-----------------------------------------------------------------------------------+
  | SHARED ENCODER: Spatiotemporal Transformer Backbone with Cross-Attention          |
  | - Ingests 4D Atmospheric/Satellite Tensor Grid                                    |
  | - Cross-Attends Dynamic Satellite Radiances with Thermodynamic IMDAA Baselines    |
  +-----------------------------------------+-----------------------------------------+
                                            |
              +-----------------------------+-----------------------------+
              |                             |                             |
              v                             v                             v
  +-----------------------+     +-----------------------+     +-----------------------+
  | HEAD 1: Thunderstorm  |     | HEAD 2: Cloudburst    |     | HEAD 3: Flash Flood   |
  | Probability Head      |     | Probability Head      |     | Prediction Head       |
  | - Convective Onset    |     | - Extreme Rain Map    |     | - Topographic Routing |
  |   Map (0-100%)        |     |   (>100 mm/hr risk)   |     |   Coupling via HAND   |
  +-----------+-----------+     +-----------+-----------+     +-----------+-----------+
              |                             |                             |
              +-----------------------------+-----------------------------+
                                            v
[LAYER 4: EXPLAINABLE AI (XAI) & POST-PROCESSING ENGINE]
  +-----------------------------------------------------------------------------------+
  | - Integrated Gradients & Cross-Attention Attribution Heatmap Generator            |
  | - Spatial Polygon Vectorizer (GeoJSON Risk Polygons)                              |
  | - OASIS CAP v1.2 Alert Formatter                                                 |
  +-----------------------------------------+-----------------------------------------+
                                            v
[LAYER 5: PRESENTATION & DISSEMINATION SERVICES]
  +------------------------------------------+----------------------------------------+
  | INTERACTIVE DASHBOARD (React + Mapbox GL)| AUTOMATED ALERT API (FastAPI REST / WS)|
  | - Dynamic DEM 3D Surface Overlays        | - CAP v1.2 Webhook Push to NDMA/SDMA   |
  | - Layer Toggles (IWV, CTT, CAPE, Risk)   | - Web Push / Mobile SMS Trigger Payload|
  | - XAI Trigger Breakdown Side-Panel       | - Emergency First Responder Geo-Fencing|
  +------------------------------------------+----------------------------------------+

Detailed Data-Flow Narrative

1. Ingestion & Preprocessing Phase:
   Every 15 minutes, automated cron workers fetch the latest INSAT-3D/3DR Imager files from MOSDAC SFTP/HTTPS mirrors [RESEARCHED]. The pipeline uses Python satpy and gdal to extract raw counts, calibrate to Brightness Temperatures (BT) for Water Vapor and TIR channels, and compute Integrated Water Vapor (IWV) proxies [RESEARCHED, INFERENCE]. Simultaneously, the system reads the corresponding IMDAA/NCUM thermodynamic grid frame (temperature, humidity, $U/V$ winds across 1000 to 200 hPa) [RESEARCHED, INFERENCE].

2. Alignment & Feature Vectorization:
   The raw multi-source rasters are resampled to a standardized $0.04^\circ \times 0.04^\circ$ spatial grid (~4 km resolution) covering the target geographic box [INFERENCE, ASSUMPTION]. Temporal derivative kernels calculate $\Delta \text{CTT}/\Delta t$ (cooling rate) and $\Delta \text{IWV}/\Delta t$ (moisture convergence rate) across the preceding 4 satellite frames ($T-60\text{m}, T-45\text{m}, T-30\text{m}, T-0\text{m}$) [INFERENCE]. Low-level horizontal wind divergence ($\nabla \cdot \vec{V}$) and bulk wind shear between 850 hPa and 200 hPa are computed [INFERENCE].

3. Multi-Task Neural Network Forward Pass:
   The aligned multi-channel tensor $[B, T=4, C=12, H, W]$ is fed into the Multi-Task Transformer [INFERENCE]. The shared encoder extracts joint spatiotemporal embeddings via 3D Convolutional Stem layers followed by Cross-Attention Transformer blocks [INFERENCE]. The representations branch into three output heads:

   • Head 1 (Thunderstorm Head): Predicts pixel-wise convective storm onset probability map for $T+2\text{h}$ to $T+6\text{h}$ [PS, INFERENCE].
   • Head 2 (Cloudburst Head): Predicts probability of extreme localized rainfall (>100 mm/hr rate) [PS, INFERENCE].
   • Head 3 (Flash Flood Head): Multiplies predicted rainfall probability by surface flow accumulation masks derived from CartoDEM (using HAND - Height Above Nearest Drainage index), yielding a topography-routed flash flood inundation risk map [PS, INFERENCE].

4. XAI Attribution & Alert Dispatch:
   If any grid cell breaches predefined threat thresholds (e.g., Cloudburst risk > 70%), the XAI module computes feature saliency using Integrated Gradients [INFERENCE]. Saliency vectors (e.g., "75% risk driven by 12°C/hr CTT drop rate + IWV spike of +18%") are attached to the risk polygon [INFERENCE]. The post-processor converts risk rasters into GeoJSON vector polygons and publishes OASIS CAP v1.2 XML messages to the API endpoint for delivery to NDMA/SDMA dashboards [PS, INFERENCE].

12. Technology Stack

13. AI/ML Justification

Detailed AI Component Analysis

+-----------------------------------------------------------------------------------+
|                        MULTI-TASK AIPredictive ARCHITECTURE                       |
+-----------------------------------------------------------------------------------+
                                  INPUT TENSOR
                    [Satellite Imagery + IMDAA Thermodynamic]
                                       |
                                       v
                    +-------------------------------------+
                    |   SHARED SPATIOTEMPORAL ENCODER     |
                    | (Cross-Attention Transformer Block) |
                    +------------------+------------------+
                                       |
           +---------------------------+---------------------------+
           |                           |                           |
           v                           v                           v
+---------------------+     +---------------------+     +---------------------+
| THUNDERSTORM HEAD   |     |   CLOUDBURST HEAD   |     |  FLASH FLOOD HEAD   |
| Binary Classifier   |     | Focal Loss Regressor|     | HAND Coupled Grid   |
| (Conv2D + Sigmoid)  |     | (>100mm/hr risk)    |     | (Topographic Mask)  |
+---------------------+     +---------------------+     +---------------------+

1. Shared Encoder Backbone:

   • Purpose: Extract unified 3D spatiotemporal atmospheric feature representations across satellite imagery and thermodynamic grids [PS].
   • Model Type: Spatiotemporal Cross-Attention Transformer (3D ConvStem + Swin/Transformer Blocks) [PS, INFERENCE].
   • Input: 5D Tensor $[B, T=4, C=12, H=256, W=256]$ (Channels: WV, TIR1, TIR2, IWV, CTT Drop, CAPE, CIN, $U_{10\text{m}}, V_{10\text{m}}$, Divergence, Vertical Shear, Elevation) [INFERENCE].
   • Output: Latent feature tensor $[B, C'=512, H/4, W/4]$ [INFERENCE].
   • Data Required: Synchronized historical INSAT-3D HDF5 files and IMDAA NetCDF reanalysis grids [PS, RESEARCHED].
   • Why AI: Atmospheric convection involves non-linear, high-dimensional fluid dynamics and thermodynamic interactions that physical NWP models take hours to compute; deep learning approximates these complex mappings in seconds [PS, KNOWN].
   • Evaluation: Spatial $F_1$-score, Critical Success Index (CSI), and Continuous Ranked Probability Score (CRPS) [INFERENCE].
   • Failure Mode & Mitigation: Missing satellite frames due to sensor outage; system falls back to linear temporal interpolation of previous frames and flags degraded confidence score in XAI panel [INFERENCE].
   • Human Oversight: Operational forecaster retains final authority to override automated red alert dispatches [INFERENCE].

2. Thunderstorm Prediction Head:

   • Purpose: Generate spatial probability map for convective storm onset 2–6 hours ahead [PS].
   • Model Type: 2D Transposed Convolutional / U-Net decoder head with Sigmoid activation [INFERENCE].
   • Input: Latent backbone feature tensor [INFERENCE].
   • Output: Pixel-wise probability grid $[0.0, 1.0]$ for severe thunderstorm presence [INFERENCE].
   • Why AI: Thunderstorms form rapidly when eroding CIN coincides with buoyant CAPE and localized low-level convergence; rules cannot capture spatial pattern evolution [PS, KNOWN].

3. Cloudburst Prediction Head:

   • Purpose: Forecast extreme localized rainfall events (>100 mm/hour over a small area) [PS, KNOWN].
   • Model Type: 2D Convolutional head trained with Focal Loss to handle extreme class imbalance [INFERENCE].
   • Input: Latent backbone feature tensor [INFERENCE].
   • Output: Probability mask of cloudburst precipitation intensity [INFERENCE].
   • Why AI: Cloudbursts are sub-grid microscale phenomena invisible to coarse NWP grids [PS, KNOWN].

4. Flash Flood Prediction Head (Hydrological Coupling):

   • Purpose: Predict ground surface inundation risk by channeling cloudburst precipitation along terrain slopes [PS].
   • Model Type: Topographic attention mask overlay multiplying predicted rainfall intensity by CartoDEM HAND (Height Above Nearest Drainage) index and D8 flow accumulation [PS, INFERENCE].
   • Input: Cloudburst Head probability grid $\times$ Static CartoDEM HAND Raster Layer [PS, INFERENCE].
   • Output: Flash flood risk probability map [PS, INFERENCE].
   • Why AI + GIS Hybrid: Atmospheric predictions provide the volume of water; static topographic elevation models channel the water. Pure AI without DEM physics ignores gravity routing; pure GIS without AI lacks precipitative lead time [PS, INFERENCE].

Non-AI Components (Where AI Should NOT Be Used)

• Unit Conversions & Radiometric Calibration: Raw satellite counts converted to Brightness Temperature using official lookup tables (LUT) must use deterministic physics formulas, NOT neural networks [INFERENCE].
• Spatial CRS Reprojections: Converting satellite lat-long projection grids to EPSG:4326 must use deterministic GDAL/Proj library transformations [INFERENCE].
• OASIS CAP Alert Generation: Parsing alert triggers into XML schemas must be done via deterministic standard template formatters to guarantee zero syntax hallucination [INFERENCE].

14. Data Architecture & Offline/Low-Network Design

Schema Architecture (PostGIS Spatial Entities)

-- 1. STATIONS & GRIDS TABLE
CREATE TABLE spatial_grids (
    grid_id SERIAL PRIMARY KEY,
    grid_code VARCHAR(32) UNIQUE NOT NULL,
    geom GEOMETRY(Polygon, 4326) NOT NULL,
    elevation_mean FLOAT,
    hand_index_mean FLOAT,
    drainage_basin_id INT
);

-- 2. LIVE INFERENCE RUNS TABLE
CREATE TABLE inference_runs (
    run_id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    timestamp_utc TIMESTAMP WITH TIME ZONE NOT NULL,
    lead_time_hours INT NOT NULL, -- 2, 4, or 6 hours
    model_version VARCHAR(16) NOT NULL,
    status VARCHAR(16) CHECK (status IN ('RUNNING', 'COMPLETED', 'FAILED'))
);

-- 3. MULTI-HAZARD PREDICTIONS TABLE
CREATE TABLE hazard_predictions (
    pred_id BIGSERIAL PRIMARY KEY,
    run_id UUID REFERENCES inference_runs(run_id),
    grid_id INT REFERENCES spatial_grids(grid_id),
    thunderstorm_prob FLOAT CHECK (thunderstorm_prob BETWEEN 0 AND 1),
    cloudburst_prob FLOAT CHECK (cloudburst_prob BETWEEN 0 AND 1),
    flash_flood_prob FLOAT CHECK (flash_flood_prob BETWEEN 0 AND 1),
    risk_level VARCHAR(8) CHECK (risk_level IN ('GREEN', 'YELLOW', 'ORANGE', 'RED')),
    xai_saliency_json JSONB -- Stores top driver attributions (e.g. {"IWV_surge": 0.45, "CTT_drop": 0.35})
);

-- Spatial index for sub-millisecond query performance
CREATE INDEX idx_hazard_pred_geom ON spatial_grids USING GIST (geom);
CREATE INDEX idx_hazard_risk_level ON hazard_predictions (risk_level);

Offline & Low-Network Resiliency Design

+-----------------------------------------------------------------------------------+
|                        OFFLINE / LOW-NETWORK SYNC ARCHITECTURE                    |
+-----------------------------------------------------------------------------------+

   +-----------------------+                    +-----------------------+
   |   CENTRAL CLOUD/HPC   |                    | FIELD RESPONDER MOBILE|
   | (Generates GeoJSON    |                    | (PWA / Android Client)|
   |  & Vector Tiles)      |                    |                       |
   +-----------+-----------+                    +-----------+-----------+
               |                                            |
               | Push Heavy Layers (Full Connectivity)      |
               |------------------------------------------->|
               |                                            |
               | Network Drops / 2G Connection             |
               | - Fallback to SMS CAP Alert Payloads       |
               | - Cache Vector Tiles in IndexedDB          |
               |<-------------------------------------------|

1. Bandwidth Reduction & Tile Vectorization:
   Raw raster probability grids (MBs per frame) are raster-to-vector converted on the server into lightweight GeoJSON polygons compressed with Mapbox Vector Tiles (mvt). This reduces payload size from ~15 MB to < 100 KB per update [INFERENCE].

2. Client-Side Progressive Web App (PWA) Caching:
   The frontend GIS dashboard uses Service Workers and IndexedDB to cache base CartoDEM elevation maps and administrative boundaries locally [INFERENCE]. When internet connectivity is lost, cached maps display the last known valid nowcast layer with a prominent "OFFLINE - DATA AS OF [TIMESTAMP]" warning banner [INFERENCE].

3. Fallback SMS / Cell Broadcast Payload Structure:
   For remote field teams with total broadband loss, the backend automatically converts high-risk alerts into compressed CAP-compliant text payloads sent via SMS / Cell Broadcast:
   [SACHET-ALERT] RED WARNING: Cloudburst & Flash Flood predicted for Valley Watershed #402 (Chamoli) between 14:00-16:00 IST. Evacuate river banks. Triggers: IWV Surge 85%, CTT Drop 14C/hr. [INFERENCE].

15. Security, Privacy & Governance

1. Sensitive & Location Data Protection:

   • Meteorological satellite and elevation data are public/open domain data sources [KNOWN].
   • User location data collected via mobile app or dashboard browser permissions must be processed purely in-memory on the client side to check against risk polygons, ensuring no citizen location tracking histories are stored on central servers [INFERENCE].

2. Authentication & Authorization (RBAC):

   • Public Users: Read-only access to published threat maps and public advisory bulletins via CDN [INFERENCE].
   • Disaster Management Officials (NDMA/SDMA): Authenticated via OAuth2 / Gov Single Sign-On (SSO) with access to raw XAI parameter modules, manual alert validation toggles, and direct emergency alert push triggers [INFERENCE].

3. Data Integrity & Audit Logging:

   • Every inference run, model weight checksum, input satellite file hash, and manual forecaster override is written to an append-only PostgreSQL audit log table [INFERENCE].
   • Ensures transparent legal accountability if an extreme event is missed or false alarms trigger costly evacuations [INFERENCE].

4. Regulatory & Scientific Validation Notice:
   As required for sensitive safety-critical disaster management systems, the output produced during the Smart India Hackathon is an academic prototype and proof-of-concept [PS, INFERENCE]. Operational deployment requires formal scientific verification and validation (V&V) by MoES, NCMRWF, and IMD evaluation committees [PS, INFERENCE].

16. MVP Scope (Hackathon Build)

+-----------------------------------------------------------------------------------+
|                            HACKATHON MVP SCOPE BOUNDARY                           |
+-----------------------------------------------------------------------------------+
|  IN SCOPE (Must Build for Live Judging)                                           |
|  - Historical Ingestion Pipeline (20+ Western Himalayan Cloudburst/Storm Events)  |
|  - Pre-trained Multi-Task Model (PyTorch -> ONNX) generating 3 Risk Maps          |
|  - Interactive Mapbox GIS Dashboard with 3D DEM Overlay                           |
|  - Functional XAI Visual Panel showing IWV, CTT, CAPE contribution bars           |
|  - Live CAP v1.2 REST API generating valid JSON/XML alerts                        |
+-----------------------------------------------------------------------------------+
|  OUT OF SCOPE / DESCRIBED ONLY (Future Production)                                |
|  - Real-time hardware integration with ISRO satellite ground station receivers     |
|  - Direct automated Cell Broadcast SMS push via C-DOT telecom towers              |
|  - Multi-node distributed HPC GPU cluster execution                               |
+-----------------------------------------------------------------------------------+

In Scope (MUST BUILD)

• Multi-Modal Data Pipeline: Automated Python script parsing pre-downloaded INSAT-3D HDF5 satellite imagery, IMDAA NetCDF thermodynamic grids, and CartoDEM GeoTIFFs for historical test cases (e.g., Himachal / Uttarakhand heavy rain events) [INFERENCE].
• Core AI Predictive Engine: Functional PyTorch multi-task spatiotemporal model exported to ONNX format, generating spatial outputs for:
  1. Thunderstorm Probability Map [PS]
  2. Cloudburst Risk Map (>100 mm/hr threshold) [PS]
  3. Flash Flood Inundation Map (HAND topographically routed) [PS]
• Interactive Web Dashboard: React + Mapbox GL JS spatial interface allowing judges to:
  • Select historical test dates or run real-time simulated scans [INFERENCE].
  • Render color-coded risk heatmaps layered on top of 3D CartoDEM terrain [PS, INFERENCE].
  • Click any grid pixel to trigger the Explainable AI (XAI) breakdown panel [PS, INFERENCE].
• Automated CAP API Service: FastAPI service generating valid OASIS CAP v1.2 XML/JSON emergency alert payloads with GeoJSON threat polygons [PS, INFERENCE].

Described Only / Out of Scope (DO NOT BUILD)

• Direct hardware-level connection to live ISRO satellite telemetry downlink dishes [INFERENCE].
• Live telecom tower Cell Broadcast integration (simulated via WebPush / Mock API response during demo) [INFERENCE].
• Multi-region global forecasting outside the Indian subcontinent [INFERENCE].

17. Live Demo Simulation Plan

To present a zero-failure, convincing live demonstration during SIH judging without relying on unpredictable live satellite feed availability:

1. Historical Event Simulator ("Replay Mode"):

   • Setup: Pre-load complete 15-minute time-series sequences from iconic extreme weather events (e.g., Himachal Pradesh July 2023 cloudburst or Uttarakhand flash flood cases) [RESEARCHED, ASSUMPTION].
   • Execution: A dashboard "Live Feed Simulator" toggle steps through historical satellite scans frame-by-frame (e.g., 10:00 AM, 10:15 AM, 10:30 AM) [INFERENCE].
   • Judges' View: Judges observe the IWV accumulation building and CTT dropping rapidly on the map screen 3 hours before the actual historical cloudburst hit [PS, INFERENCE].

2. Interactive Parameter Injector ("What-If Scenario"):

   • Setup: A hidden simulation slider panel allowing judges to artificially boost IWV moisture values or steepen CTT drop rates [INFERENCE].
   • Execution: Slide IWV rate from normal to +25% accumulation [INFERENCE].
   • Judges' View: The multi-task AI engine instantly re-indexes the grid in < 2 seconds, upgrading the flash flood head output from "YELLOW" to "RED", and triggering an automated CAP alert popup on screen [PS, INFERENCE].

3. Simulated Alert Delivery Journey:

   • Show live API request payload generated by FastAPI (/api/v1/alerts/latest) formatted in strict OASIS CAP v1.2 XML format [INFERENCE].
   • Trigger a live WebPush notification on a test smartphone placed on the judging table [INFERENCE].

Honesty Disclaimer: The team will explicitly inform judges that satellite inputs are replayed from archived MOSDAC HDF5 files to guarantee demo stability [INFERENCE].

18. Implementation Roadmap

+-----------------------------------------------------------------------------------+
|                        IMPLEMENTATION ROADMAP (Target: 30 Sept 2026)              |
+-----------------------------------------------------------------------------------+
 Phase 1: Data Curation & Preprocessing (Month 1 - 2)
 [Dataset assembly: INSAT-3D HDF5, IMDAA NetCDF, CartoDEM, Event Catalog]
                        |
                        v
 Phase 2: Model Architecture & Training (Month 3 - 4)
 [PyTorch Spatiotemporal Cross-Attention Transformer + MTL Heads]
                        |
                        v
 Phase 3: Hydrological Routing & XAI Integration (Month 5)
 [CartoDEM HAND index coupling + Integrated Gradients module]
                        |
                        v
 Phase 4: API & Web GIS Dashboard Development (Month 6)
 [FastAPI backend + Mapbox 3D terrain frontend + CAP API]
                        |
                        v
 Phase 5: Testing, Optimization & Hackathon Polishing (Month 7 / Final Lead-up)
 [ONNX acceleration, historical event validation, demo dry-runs]
+-----------------------------------------------------------------------------------+

Phase Details & Milestones

• Phase 1: Data Ingestion & Preprocessing Pipeline (Weeks 1–6)

  • Milestone 1.1: Download and organize historical INSAT-3D/3DR satellite files for 25 severe weather events across J&K, Himachal Pradesh, and Uttarakhand [RESEARCHED, ASSUMPTION].
  • Milestone 1.2: Build spatial reprojection and interpolation scripts converting IMDAA, INSAT, and CartoDEM into a unified $0.04^\circ$ spatial grid tensor [INFERENCE].

• Phase 2: Multi-Task AI Engine Development (Weeks 7–14)

  • Milestone 2.1: Construct PyTorch 3D Convolutional + Cross-Attention backbone [INFERENCE].
  • Milestone 2.2: Implement 3 output prediction heads (Thunderstorm, Cloudburst, Flash Flood) with dynamic multi-task loss weighting [PS, INFERENCE].
  • Milestone 2.3: Train model on historical benchmark dataset; optimize $F_1$-score and export backbone to ONNX format [INFERENCE].

• Phase 3: Topographic Coupling & XAI Saliency (Weeks 15–18)

  • Milestone 3.1: Compute CartoDEM D8 flow accumulation and HAND rasters using RichDEM [INFERENCE].
  • Milestone 3.2: Integrate PyTorch Captum / Integrated Gradients module for pixel-level feature attribution calculation [INFERENCE].

• Phase 4: Dashboard & Alert API Development (Weeks 19–23)

  • Milestone 4.1: Develop FastAPI REST services for inference serving and CAP v1.2 alert generation [INFERENCE].
  • Milestone 4.2: Build React + Mapbox GL JS frontend rendering 3D terrain, risk overlays, and XAI trigger panels [INFERENCE].

• Phase 5: Validation, Benchmarking & Demo Polish (Weeks 24–26)

  • Milestone 5.1: Benchmark end-to-end pipeline latency (< 15 seconds forward pass) [ASSUMPTION].
  • Milestone 5.2: Perform dry-run live demo scenario rehearsals ahead of the 30 September 2026 submission deadline [PS, ASSUMPTION].

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon Prototype Cost (Minimal / Free Tier)

Production Deployment Cost (Government Scale)

22. Differentiation

23. Future Scope

1. Integration with Doppler Weather Radar (DWR) Radial Velocity Data: Incorporate real-time IMD DWR reflectivity ($Z$) and radial velocity ($V$) volumes to further refine sub-kilometer convective cell tracking as radar networks expand [RESEARCHED, INFERENCE].
2. Crowdsourced Public Weather Observations: Ingest geo-tagged citizen photos and reports (via IMD Public Observation portal integration) to provide ground-truth validation for active cloudburst inundation [RESEARCHED, INFERENCE].
3. On-Device Edge Inference for First Responders: Quantize the multi-task model to INT8 precision for lightweight execution on ruggedized field tablets, enabling local offline NOWCAST updates even when remote communications are totally severed [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions Log

• [ASSUMPTION 01]: Model forward pass execution under 15 seconds per grid is achievable on standard T4/RTX GPUs using ONNX Runtime quantization.
• [ASSUMPTION 02]: Historical training can be effectively conducted using 20+ documented Western Himalayan extreme rain events (2010–2024) compiled from IMD and public disaster archives.
• [ASSUMPTION 03]: Replaying pre-downloaded MOSDAC HDF5 satellite files is accepted by SIH judges as a reliable proxy for live ISRO satellite feeds.
• [ASSUMPTION 04]: Resampling satellite Water Vapor channels (8 km) to a unified 0.04° (~4 km) spatial grid preserves sufficient spatial gradient variance for IWV tracking.
• [ASSUMPTION 05]: Total hackathon execution software tools (PyTorch, FastAPI, Mapbox, PostGIS, RichDEM) are 100% open-source or offer sufficient free tier access.

Consolidated Open Items & Validation Checklist

• [NEEDS VALIDATION 01]: Direct real-time API subscription key availability for MOSDAC satellite feeds during the hackathon phase vs. relying on manual FTP/HTTPS downloads.
• [NEEDS VALIDATION 02]: Exact operational update latency of IMDAA/NCUM analysis baselines at NCMRWF during live operations.
• [NEEDS VALIDATION 03]: Formal integration guidelines and endpoint credentials for pushing test CAP alerts into the NDMA SACHET sandbox environment.

Final Recommendation

Build the Multi-Task Spatiotemporal Cross-Attention Transformer Engine with Hydrological Topographic Coupling [PS, INFERENCE]. This architecture directly satisfies every requirement of SIH26077, bypasses NWP latency, fuses satellite and thermodynamic inputs, and delivers actionable, explainable 2-to-6-hour warnings [PS].

Recommended Stack

• Frontend: React.js, Mapbox GL JS, Tailwind CSS [INFERENCE]
• Backend Services: FastAPI (Python 3.12), PostGIS / PostgreSQL 16 [INFERENCE]
• AI / ML Pipeline: PyTorch 2.x, PyTorch Lightning, ONNX Runtime, Captum (XAI) [INFERENCE]
• Hydrology & Data Processing: Satpy, h5py, Xarray, RichDEM, GDAL [RESEARCHED, INFERENCE]

Key Summary Metrics

• Biggest Dependency: Access to clean, historical INSAT-3D/3DR satellite HDF5 files from MOSDAC for extreme convective events [RESEARCHED, INFERENCE].
• Biggest Risk: Extreme class imbalance (cloudbursts are rare events), leading to potential model under-prediction without careful loss weighting [KNOWN, INFERENCE].
• Biggest Technical Challenge: Aligning multi-scale inputs (8 km satellite WV, 12 km IMDAA thermodynamic grids, and 30 m CartoDEM rasters) into a synchronized 5D tensor grid [PS, INFERENCE].

First 5 Actions for Student Team

1. Register on MOSDAC (mosdac.gov.in) and NCMRWF RDS (rds.ncmrwf.gov.in) to download sample INSAT-3D HDF5 and IMDAA NetCDF files for 5 historical cloudburst dates [RESEARCHED].
2. Set up a shared GitHub repository with Python satpy, xarray, h5py, and gdal to test satellite raster extraction and IWV/CTT calculation [INFERENCE].
3. Build a spatial grid resampling script to convert satellite and reanalysis data into a unified $0.04^\circ$ spatial grid tensor [INFERENCE].
4. Implement a baseline PyTorch Multi-Task U-Net / ConvLSTM network with three distinct output heads [INFERENCE].
5. Create a Mapbox GL JS 3D terrain frontend prototype rendering sample vector GeoJSON risk polygons over CartoDEM elevation [INFERENCE].

Validation Before Coding

• Verify that downloaded INSAT-3D HDF5 files contain uncorrupted Water Vapor (6.5–7.1 $\mu$m) and TIR-1 (10.2–11.2 $\mu$m) channel arrays [RESEARCHED].
• Test that RichDEM / WhiteboxTools successfully generates flow accumulation and HAND rasters from downloaded SRTM/CartoDEM GeoTIFFs without memory overflow [INFERENCE].

Quantitative Evaluation

• Buildability Score: 88/100 (Highly feasible for a 4–6 student team using PyTorch, ONNX, Mapbox, and pre-processed sample datasets within the available timeline) [INFERENCE].
• Confidence: HIGH (All explicit requirements mapped directly to established software libraries, proven deep learning patterns, and official open datasets) [PS, RESEARCHED, INFERENCE].

Research Sources

1. Title: IMDAA: High-Resolution Satellite-Era Reanalysis for the Indian Monsoon Region
   URL: https://journals.ametsoc.org/view/journals/clim/34/12/JCLI-D-20-0412.1.xml
   Source Type: Official Academic / Journal Publication (AMS Journals)
   Supported Claim: Confirms IMDAA is a 12 km high-resolution regional atmospheric reanalysis dataset over India produced by NCMRWF, IMD, and UK Met Office, available via NCMRWF RDS (rds.ncmrwf.gov.in).
   Reliability Assessment: High (Peer-reviewed scientific publication).

2. Title: INSAT-3D / 3DR Data Products & Payload Specifications - MOSDAC
   URL: https://www.mosdac.gov.in/insat-3d-data-products
   Source Type: Official Government Data Portal Documentation (ISRO / SAC MOSDAC)
   Supported Claim: Confirms INSAT-3D/3DR payload channel specifications: Water Vapor (WV, 8 km), Thermal Infrared (TIR1/TIR2, 4 km), QPE products, and HDF5 file formats available via MOSDAC.
   Reliability Assessment: High (Official Government / Space Agency Documentation).

3. Title: South Asia Flash Flood Guidance System (SAsiaFFGS)
   URL: https://reliefweb.int/report/india/south-asia-flash-flood-guidance-system-launched
   Source Type: Official International / Government Press Release (WMO / IMD / PIB)
   Supported Claim: Documents the operational status of SAsiaFFGS launched by IMD/WMO for watershed-level flash flood guidance across South Asia.
   Reliability Assessment: High (Official WMO / Government Source).

4. Title: SACHET - National Disaster Alert Portal & Common Alerting Protocol
   URL: https://sachet.ndma.gov.in/
   Source Type: Official Government Portal (NDMA / C-DOT)
   Supported Claim: Verifies NDMA's SACHET platform as India's official CAP-compliant (OASIS CAP v1.2) multi-lingual geo-targeted disaster alert dissemination hub.
   Reliability Assessment: High (Official Government Portal).

5. Title: BharatBench: Benchmark Dataset derived from IMDAA Reanalysis for ML Weather Forecasting
   URL: https://github.com/MASLABnitrkl/BharatBench
   Source Type: Open Source Research Repository / Benchmark Dataset
   Supported Claim: Demonstrates practical usage of IMDAA reanalysis grids in deep learning (ConvLSTM/CNN) weather prediction frameworks tailored for the Indian subcontinent.
   Reliability Assessment: High (Open Source Academic Research Codebase).