"From Outer Space to the Courtroom — In Real Time."
Autonomous satellite-powered compliance monitoring that detects, classifies, and legally documents unauthorized land-use changes using a hybrid Siamese-SNN, LangGraph agents, and blockchain evidence.
The Problem
Urban violations are invisible to the human eye. Manual inspections cover <2% of city area per year. By the time an officer arrives, the building is 4 floors high.
The Solution
What if the inspector never left the office? GeoGuard AI autonomously detects, classifies, and legally documents every violation.
The Pipeline
Fully autonomous 9-step pipeline. No human in the loop.
User draws a bounding box on the Leaflet.js map to define the Area of Interest.
Authenticates with ESA Copernicus CDSE. Downloads Sentinel-2 bands B02, B03, B04, B08 at 10m/pixel.
SCL filtering + morphological dilation (50m buffer). All cloud zones hard-overridden to zero.
Sub-pixel co-registration ensures Before/After stacks are perfectly aligned.
Siamese-SNN processes both dates through shared U-Net encoder. |F_A − F_B| difference map computed over 8 spiking timesteps.
ΔNDVI, ΔNDBI, ΔMNDWI indices classify changes as construction, vegetation loss, or water encroachment.
Rule engine evaluates against Gujarat GDCR 2017, SUDA 2035, Tapi Riverfront regulations. Risk scored 0–100.
Images, PDF report, metadata uploaded. SHA-256 evidence hash sealed in database.
Structured WhatsApp message dispatched with TinyURL links via WAHA Docker API.
The AI Brain
Not just a CNN. A brain that spikes. First known application of Spiking Neural Networks to satellite change detection.
Spectral Analysis
Identifying materials from space using normalized difference indices.
Agentic AI
LangGraph state machine • Llama 3.3 70B via OpenRouter • 8 autonomous tools • real data from Supabase.
Evidence & Alerting
WAHA Docker API → TinyURL shortener → structured message to field officer. Also triggerable from the chatbot.
"Even the most powerful official cannot change what the satellite saw."
Model Training
Trained on OSCD (Onera Satellite Change Detection) — 24 cities, expert-labeled change masks.
# 1. Download dataset python scripts/download_oscd.py # 2. Reorganize python scripts/reorganize_oscd.py # 3. Train (Colab notebook) # Train_SiameseSNN_Colab_v3.ipynb class SiameseSNN(nn.Module): def __init__(self, in_ch=4): self.encoder = UNetEncoder(in_ch) self.decoder = SNNDecoder( time_steps=8, beta=0.85 ) def forward(self, a, b): f_a = self.encoder(a) f_b = self.encoder(b) diff = torch.abs(f_a - f_b) return self.decoder(diff) # Loss: Dice(1.0) + BCE(0.5) + Focal(0.5) # Output: data/models/siamese_snn_oscd.pth
Technology
Production-grade. 100% open source. Total license cost: ₹0.
Impact
Get Started
Clone → configure → scan. Takes about 5 minutes.
git clone https://github.com/shaktisingh5580/PFL-SNN-BACKEND.git
cd PFL-SNN-BACKENDPython 3.10+ required.
pip install -r requirements.txtcp .env.example .env
| SH_CLIENT_ID | Copernicus CDSE Client ID |
| SH_CLIENT_SECRET | SentinelHub Secret |
| SUPABASE_URL | Supabase project URL |
| SUPABASE_KEY | Service role key |
| OPENROUTER_API_KEY | For Llama 3.3 70B |
Place pre-trained weights or train via Colab.
mkdir -p data/models
# Place siamese_snn_oscd.pth herepython scripts/scan_vesu_surat.pypython start_server.py # http://0.0.0.0:8000docker-compose up --buildFuture