Multimodal Earth Observation Analysis & Visual Grounding Workbench
Developed for Smart India Hackathon (SIH26167 / ISRO Earth Observation Challenge)
SATYA (Satellite Analysis Through Your Assistant) is an agentic Earth Observation (EO) and geospatial intelligence workbench. It translates natural language questions into deterministic remote-sensing execution pipelines — orchestrating specialist computer vision models, calculating sub-pixel geospatial measurements, performing optical-SAR cross-corroboration, and producing grounded audit dossiers.
Unlike generic conversational models that hallucinate spatial masks and coordinates, SATYA employs a Controlled Agentic Architecture:
- Natural language intent routing to deterministic remote-sensing models.
- Direct pixel-to-ground area calculations using UTM coordinate reference systems and affine geotransforms.
- Multi-sensor verification fusing Sentinel-2 multispectral optical reflectance with Sentinel-1 C-band Synthetic Aperture Radar (SAR) polarimetric backscatter.
- Isotonic temperature-scaled confidence calibration with automatic false-change rejection.
- Visual Question Answering & Sub-Pixel Grounding: Point-and-click polygon grounding, bounding vectors, and class segmentation across optical scenes.
-
Bi-Temporal Differential Change Detection: Siamese feature extraction, NDVI differencing, and metric area change estimation in hectares (
$ha$ ) and$km^2$ . -
Optical + SAR Multi-Sensor Corroboration: Cross-verifies optical surface reflectance with Sentinel-1 C-band SAR polarimetric backscatter (
$+11.4\text{ dB}$ double-bounce confirmation) to eliminate false alarms from cloud shadows or seasonal variations. - Interactive Geospatial Stage: Split-swipe bi-temporal comparative inspection, multispectral False Color (NIR) rendering, and real-time polygon inspection tooltips.
- Controlled Specialist Tool Registry: Complete schema definitions, latency budgets, and failure thresholds for deep learning models.
- SATYA-BENCH Evaluation Suite: Automated benchmark matrix evaluating spatial IoU, expected calibration error (ECE), and false change rejection.
- Exportable PDF Dossiers: Download comprehensive geospatial intelligence audit reports with single-click PDF generation.
- Custom AOI Ingestion: Support for uploading and analyzing custom multispectral and SAR GeoTIFF/HDF5 rasters.
- Frontend: React 18, TypeScript, Tailwind CSS, Lucide Icons, jsPDF.
- Backend / Serverless: Node.js, Express, Vercel Serverless Functions (
/api/*). - Orchestration / LLM: Gemini 2.5 (
@google/genaiSDK) with fallback to deterministic local geospatial pipelines. - Deployment Targets: Vercel, Cloud Run, Docker.
- Node.js 18+ or Bun
- npm or yarn
# Clone the repository
git clone https://github.com/Aditya20-y/SATYA.git
cd SATYA
# Install dependencies
npm install# Start the development server (port 3000)
npm run devOpen http://localhost:3000 in your browser.
SATYA is pre-configured with vercel.json and a serverless API handler (/api/index.ts).
- Push your repository to GitHub:
https://github.com/Aditya20-y/SATYA - Go to vercel.com/new and import
SATYA. - Set the following build settings:
- Framework Preset:
Vite - Build Command:
vite build - Output Directory:
dist
- Framework Preset:
- (Optional) Add your
GEMINI_API_KEYunder Environment Variables. - Click Deploy.
npm i -g vercel
vercel| Metric | Target / Benchmark Result | Description |
|---|---|---|
| Overall Accuracy | 94.2% | Verification across 40+ defense & disaster scenarios |
| Mean IoU (Grounding) | 0.842 | Polygon intersection over union vs ground-truth rasters |
| False Change Rejection | 96.8% | SAR dielectric filtering against seasonal vegetation drops |
| Calibration Error (ECE) | 3.8% | Isotonic temperature-scaled confidence alignment |
| Average Pipeline Latency | 420 ms | End-to-end multi-model execution and synthesis |
This project is licensed under the Apache-2.0 License.