DUPAHAR.AI :- India's Unified Civic Intelligence Platform

🌧️ DUPAHAR.AI — From Drone Data to Flood-Resilient Cities

Inspiration

Every monsoon, Indian cities drown. Not because we lack data — but because we ignore it.

India spent ₹566 Crore under the SVAMITVA scheme to drone-map 3.28 lakh villages at 3–5 cm precision. That's centimeter-level terrain intelligence covering nearly every rural settlement in the country. Yet 0% of it has ever been used for hydrology or drainage planning.

Meanwhile, ₹1.4 lakh crore in sanitation infrastructure gets built every year without slope validation — drains that don't drain, roads that flood, Abadi settlements with lanes under 2 metres wide that become completely inaccessible during monsoons. 58% of Indian citizens report severe waterlogging in their districts (LocalCircles, 2023).

The data to fix this exists. It's just locked in property records.

That's what inspired DUPAHAR.AI — not to build something new from scratch, but to unlock what India already paid for.

What It Does

DUPAHAR.AI is India's Unified Civic Intelligence Platform — a three-layer AI system that converts static SVAMITVA drone orthophotos into actionable flood prevention intelligence, automated engineering reports, and carbon credit revenue. No new surveys. No new budget. Only an intelligence upgrade.

🟢 Layer 1 — SEE (Geo-AI Feature Extraction)

A DeepLabV3+ with ResNet50 backbone, purpose-built for SVAMITVA drone orthophotos. Trained over 118 epochs on the full SVAMITVA dataset with 20x weighted oversampling for rare infrastructure classes.

91.93% Mean Validation IoU
97% Waterbody Detection Accuracy
89% RCC Building Detection

Extracts buildings (RCC, tiled, tin), roads (paved & unpaved), water bodies, and critical infrastructure — outputting QGIS-ready .shp shapefiles.

🟣 Layer 2 — SIMULATE (Physics-Informed Flood Prediction)

DUALFloodGNN — a physics-informed Graph Neural Network surrogate trained on SWMM ground-truth simulations. It respects mass conservation and forces water through alleys, not walls.

Key innovations:

Sonata-MAE Terrain Synthesis — bare-earth DTM reconstruction beneath dense village clutter, eliminating the Geometric Shortcut problem
Topological Graph Cut — building boundaries assigned zero conductivity for physically accurate flow routing
SWMM-Calibrated GNN — 10×–100× faster than full numerical solvers
4D Digital Twin — traffic-light risk corridors (Flooded / At-Risk / Safe) for real-time MCD emergency response

Validated on the Dariyapur SVAMITVA dataset (500,000 points):

Metric	Value
Avg. Training Error (MAE)	3.3 cm
Best Validation Error	1.3 mm
Binary Flood Detection Accuracy	93.78%
True Flood Recognition Rate	87.83%
MAE Reduction over baseline	72.9%

🟡 Layer 3 — SUSTAIN (Carbon MRV via TERRA.AI)

The drainage intelligence DUPAHAR.AI generates has a powerful byproduct:

Better drainage → reduced waterlogging → measurable methane reduction → high-integrity carbon credits

TERRA.AI automates the entire Measurement, Reporting & Verification (MRV) pipeline — VM0042 compliant, Verra-approved. No additional field surveys required. The voluntary carbon market is projected at ₹54,432 Crore by 2034.

Complete Pipeline Performance

Output	Time
Complete drainage intelligence	45 minutes
Manual equivalent	6–8 months
Auto-generated DPR report	5 seconds

How We Built It

The core technical stack:

Segmentation Model: DeepLabV3+ (ResNet50 backbone) trained on SVAMITVA ortho-tiles — custom loss weighting for class imbalance (buildings vs. rare infrastructure)
Terrain Synthesis: Sonata-MAE masked autoencoder for bare-earth DTM reconstruction from cluttered DSM data
Flood Model: DUALFloodGNN — a dual-branch GNN with physics constraints derived from EPA SWMM simulation outputs
Carbon Pipeline: TERRA.AI SOC (Soil Organic Carbon) mapping using multispectral drone data, validated on a 280ha test area
GIS Output: QGIS-compatible .shp shapefiles for direct municipal integration
Automation: Auto DPR (Detailed Project Report) generator — produces engineering drainage plans in under 5 seconds

We validated everything on real SVAMITVA data from Dariyapur village, not synthetic datasets.

Challenges We Faced

1. The Geometric Shortcut Problem Standard flood models trained on DSM data "cheat" — they treat building rooftops as terrain. Water appears to flow through walls rather than around them, making predictions physically impossible. We solved this with our Sonata-MAE terrain synthesis module that reconstructs bare-earth DTM beneath building footprints.

2. Class Imbalance in SVAMITVA Data Transformers, wells, and small infrastructure are rare in drone imagery but critical for drainage modeling. We applied 20x weighted oversampling for rare classes and a Mixture-of-Experts (MoE) specialist model achieving 77% recall on infrastructure detection.

3. Speed vs. Accuracy Tradeoff Full SWMM numerical solvers take days per village. We built DUALFloodGNN as a calibrated surrogate — physics-constrained but 10×–100× faster — enabling iterative drainage design in seconds.

4. Carbon MRV Without Field Surveys Traditional soil carbon MRV requires expensive on-ground sampling. We demonstrated that drone-derived spectral and terrain data, combined with the VM0042 methodology, enables fully digital MRV — validated on a real 280ha test area.

What We Learned

Government data at scale is India's most underutilized AI asset
Physics constraints in neural networks aren't optional for safety-critical applications — they're the difference between a demo and a deployable system
The path to climate resilience in India runs through drainage — and drainage starts with terrain intelligence

Impact Numbers

✅ Flood hotspots detected before water accumulates
✅ Drainage DPR generated in 45 minutes (vs. 6–8 months)
✅ Emergency response time reduced by 60–70%
✅ Zero new surveys required
✅ 85% software margin at scale

Business Model

Tier	Offering
Government Consulting	Drainage DPRs for municipal corporations
Hydrology SaaS/API	Flood intelligence APIs for Smart Cities, insurance, agri-tech
Carbon Monetization	Revenue share via TERRA.AI MRV system