PITCH HOOK
The Hook & Core Performance Numbers
[ 2D Satellite Screenshot ] ──▶ < 15 SECONDS ──▶ [ Metric 3D Twin + Voice AI ]
"TerraScope AI turns a flat 2D satellite screenshot into a metric 3D digital twin in under 15 seconds—giving first responders voice-directed building heights and flood intelligence without waiting days for drones or LiDAR."
Why “interactive twin”: it is a live, navigable 3D digital representation of a real place that users can inspect, measure, query, and use for scenario analysis.
Core Performance Snapshot
Turnaround Latency
TerraScope AI: <15 seconds
Industry Standard: 24–72 hours (Drones/LiDAR)
Impact: >5,000× faster
Height Accuracy (RMSE)
TerraScope AI: 1.45 m (calibrated)
Industry Standard: Unmeasured / Flat
Impact: Sub-story precision
Slum Building Separation
TerraScope AI: 1–3 px (0.35 m alleys)
Industry Standard: Fused into single blob
Impact: +358% structures detected
Hardware Required
TerraScope AI: Zero (ordinary web screenshot)
Industry Standard: $100k+ airborne sensors
Impact: Deployable anywhere
PROJECT DETAILS
Hackathon
Hackyard 2026
Team Name
unknown coders
Team Code
B26-0614
Group Members
Devaditta Patra; Himarghya Das
Problem & Product Comparison
Why Existing Geospatial Products Fail in Crises
Flat 2D maps hide volumetric reality: responders cannot see pancake collapses, rubble volumes (m³), or flood accumulation chutes.
TWI (Topographic Wetness Index) estimates where water is likely to accumulate based on terrain and natural surface runoffs.
Competitive Breakdown
# | Comparison Dimension | Traditional Giants (Nearmap, Ecopia, RMS, MS) | TerraScope AI |
1 | Speed to Action | Months to Weeks. Pre-scheduled airplane flights or static data purchases. Heavy hydrodynamic simulation runs. | < 15 Seconds (On-Demand). Upload any fresh satellite or drone screenshot. Instant interactive 3D digital twin generated live. |
2 | Geographic Reach | Western Suburbs Only. Nearmap/CAPE fly planes only over the US, UK, and Australia. Zero coverage in developing nations or disaster zones. | 100% Global Coverage. Works anywhere on Earth with an optical satellite pass. Zero dependency on local sensors or flight permits. |
3 | Dense Urban Resolution | Fused "Monolithic Blobs". Standard ML merges dense informal settlements and slums into single giant shapes. | 1-Meter Alley Separation. Specially trained alley-cutting algorithms isolate individual homes down to 1–3 pixel gaps. |
4 | Risk Intelligence Level | Macro-Grids & Flat Scores. 30m – 100m regional flood heatmaps or flat PDF property scores. | Micro-Building Physics + Voice. Pinpoints the exact house threatened by flood/landslide with hands-free voice intelligence. |
PIPELINE
Simple Workflow & Tech Architecture
A. The 4-Step User Workflow
01
1. Ingest
Upload a 2D satellite screenshot.
02
2. Reconstruct
Reconstruct a metric 3D twin.
03
3. Analyze
Inspect, measure, and query buildings and hazards.
04
4. Direct
Route, investigate, and narrate results by voice.
B. System Technological Architecture
2D Satellite Screenshot ⟶ ResNet50 U-Net Height/Landcover ⟶ 3D Geometric Extraction ⟶ Hazard/Damage Engines ⟶ Three.js 3D Twin + Agent37 + ElevenLabs
Satellite Ingestion
Plain 2D map screenshot input
ResNet50 U-Net
height_v2.pt multi-scale inference
3D Geometry
OpenCV + Shapely polygon & valley extraction
Hazard Engines
Flood TWI & Slope Factor-of-Safety solvers
Interactive Twin
Three.js canvas + Agent37 + ElevenLabs TTS
03
EMPIRICAL VERIFICATION
Neural Height Accuracy & Verified Benchmarks
1. Trained Neural Model Accuracy (height_v2.pt)
Focus Area | Metric | Result | Benchmark Dataset & Source |
1. Height Accuracy (Calibrated) | Height RMSE | 1.45 m | Copernicus Global DEM (GLO-30) (ESA / Airbus) |
2. Height Accuracy (Raw Pixel) | All-Pixel MAE | 1.84 m | IEEE GRSS DFC2019 Track 1 (Jacksonville / Omaha LiDAR) |
3. Building Footprint Accuracy | Building IoU | 0.775 (77.5%) | IEEE GRSS DFC2019 Urban Benchmark |
4. Building Detection Quality | F1-Score | 83.0% | IEEE GRSS DFC2019 Urban Benchmark |
2. High-Density Building Separation Benchmarks
Tested on real-world satellite images containing dense informal settlements:
China Urban Village (China Image.jpeg)
Standard Model: 193 buildings ⟶ TerraScope AI: 884 houses (+358%)
Separates tightly packed roofs across narrow 1–3 px alleyways without blob fusion.
Urban Slum District (CHINA IMAGE.png)
Standard Model: 293 buildings ⟶ TerraScope AI: 1,191 houses (+306%)
Maintains distinct individual structures across informal settlements.
Industrial Hub (India Image.png)
Preserved 592 structures with 0% false fragmentation on complex L-shaped roofs.
Avoids erroneous over-splitting while retaining accurate building outlines.
04
Multimodal Synergy: ElevenLabs + Agent37
The Zero-Hallucination Separation Principle
PyTorch (height_v2.pt) ⟶ Computes Metric Heights ⟶ OpenCV + Shapely ⟶ Extracts Geometry ⟶ Agent37 ⟶ Reasoning Tools ⟶ ElevenLabs ⟶ Spoken Audio
ElevenLabs Integration
Frontline officers wearing protective gloves/PPE tap any building or speak via microphone to receive instant spoken height, area, and hazard metrics.
Agent37 Integration
Orchestrates multi-step disaster investigations (e.g., "Compare M6.5 vs M7.2 earthquake scenarios"), persists investigation records under durable Sector IDs, and never hallucinates building dimensions.
05
CLOSING & ROADMAP
Market Potential & Live Demo Closing
1. Target Adopters
NDRF / FEMA / UN-OCHA
Instant sub-minute tactical 3D twins during active cyclone landfall.
Municipal Slum Authorities
Automated census, building counts, and drainage mapping for 1B+ informal settlers.
Telecom Recovery Cells
3D line-of-sight ray-marching to place emergency 5G cell-towers on intact roofs.
2. Immediate Roadmap
1
Bi-Temporal Siamese Networks
Train dedicated change-detection on full xBD disaster dataset.
2
Regional Indian Languages
Expand ElevenLabs synthesis into Hindi, Malayalam, and Tamil.
3
Live GIS Streams
Ingest OpenStreetMap Overpass and live USGS water gauge feeds.
4
New Business Opportunity
Selling detailed, Blender-editable 3D models of cities for game developers.
Projected market expansion, 2024–2030 (24.8% CAGR)
Target Industry Demand Share
06