NSGIS — Neuro-Symbolic Geospatial Inference System

First TNFD-aligned nature risk prototype over Agbogbloshie, Greater Accra, Ghana
Real Sentinel-2 L2A satellite data · 10m resolution · 90,000 cells · February 2026

---

What This Is

NSGIS is a working prototype that combines neural satellite image analysis with symbolic industrial ecology rules to identify informal industrial activity and generate TNFD LEAP-aligned nature risk disclosures — without requiring any pre-labeled ground truth data.

This is the first system to produce a 10-meter resolution nature risk heatmap over an informal industrial cluster (Agbogbloshie e-waste site, Greater Accra) from real satellite data.

What It Produces

• Activity classification map (7 informal industrial classes: Tannery, Textile Dyeing, E-Waste, Metal/Battery Recycling, Chemical Storage, Food Processing, Non-Industrial)
• TNFD Confidence Tier Map (Tier 1–4, aligned with TNFD LEAP framework)
• Nature Risk Heatmap at 10m resolution
• JSON TNFD disclosure report with impact driver profiles

Phase 1 Results (Real Sentinel-2 Data)

• Scene: 300×300 cells = 3km × 3km over Agbogbloshie, Accra
• Clear cells analyzed: 89,996 of 90,000 (100% after SCL cloud masking)
• Neuro-Symbolic F1 Macro: 0.485 vs Baseline Neural: 0.395 (+23%)
• Expected Calibration Error: 0.022 (target: <0.08)
• Tier 1 attribution: 20% of scene (17,972 cells)
• High-risk clusters flagged: 3,363 cells (risk score >0.75)

Quick Start

1. Install dependencies

pip install rasterio scipy scikit-learn matplotlib numpy

2. Download real Sentinel-2 data

Go to: https://browser.dataspace.copernicus.eu
Search for tile: T30NZM
Product: S2MSI2A (Level-2A atmospherically corrected)
Date: Any cloud-free scene over Accra
Download the ~1GB .SAFE file

3. Update the data path

Open data/real_data_loader.py
Change line 14 to your downloaded .SAFE folder path:

SAFE_PATH = r"C:\your\path\to\file.SAFE\file.SAFE"

4. Run the pipeline

cd nsgis
python pipeline/run_pipeline_real.py
python viz/visualize.py

All outputs saved to nsgis/outputs/

Repository Structure

nsgis/
├── data/
│   ├── sentinel_simulator.py    # Synthetic data generator
│   └── real_data_loader.py      # Real Sentinel-2 loader ← start here
├── iakg/
│   └── knowledge_graph.py       # Industrial Activity Knowledge Graph
│                                  (10 rules, 7 classes, 3-valued logic)
├── neural/
│   └── model.py                 # Neural module + NS fusion system
├── pipeline/
│   ├── run_pipeline.py          # Simulation pipeline
│   └── run_pipeline_real.py     # Real data pipeline ← run this
├── viz/
│   └── visualize.py             # All 6 publication figures
└── outputs/                     # Generated figures and reports

System Architecture

Sentinel-2 + ECOSTRESS + Sentinel-1
→ Multi-sensor fusion (cross-attention)
→ Neural perception (Clay backbone)
→ [IAKG symbolic rules] + [DeepProbLog reasoner]
→ Neuro-symbolic fusion (adaptive weighting)
→ TNFD Report + Risk Heatmap

Papers

Preprint 1 (Framework):
Gupta, R. (2025). Neuro-Symbolic Geospatial Intelligence: A Framework for Understanding Nature-Related Risks in the Informal Global South.
https://www.authorea.com/users/1001242/articles/1396647-neuro-symbolic-geospatial-intelligence-a-framework-for-understanding-nature-related-risks-in-the-informal-global-south

Preprint 2 (Full Architecture + Phase 1 Results):
Gupta, R. (2026). Operationalizing Neuro-Symbolic Geospatial Intelligence.
https://www.authorea.com/users/1001242/articles/1396039-operationalizing-neuro-symbolic-geospatial-intelligence

Phase 2 (In Progress)

• Real ECOSTRESS thermal data (NASA AppEEARS — pending)
• Real Sentinel-1 SAR data (Copernicus — pending)
• Ground truth field survey via ISODEC partnership (Ghana)
• Target journal: Nature or Remote Sensing of Environment

License

MIT License — free to use, modify, and distribute with attribution.

Contact

Rishaank Gupta · site.rishaank@gmail.com