Satellite Imagery Analysis

What we monitor

34 military sites across Russia, Belarus, and Kaliningrad using free Sentinel-2 (optical, 10m resolution) and Sentinel-1 (SAR radar) imagery from the Copernicus programme.

Data source

• Google Earth Engine — Sentinel-2 SR Harmonized + Sentinel-1 GRD
• GCS storage — thumbnails stored in gs://estwarden-satellite/
• Collection frequency: depends on satellite revisit (~5 days per site)
• Collector: satellite_collector.py

What we measure

Optical (Sentinel-2)

• RGB thumbnail — true-color image for visual comparison
• BSI (Bare Soil Index) — detects exposed ground, construction activity
• NDBI (Normalized Difference Built-up Index) — detects buildings, hard surfaces
• Fuel percentage — spectral proxy for fuel/petroleum storage activity
• Metal proxy — highlights metallic objects (vehicles, equipment)

SAR (Sentinel-1)

• Mean change (dB) — backscatter difference vs 30-day baseline
• Significant change (>3 dB) suggests vehicle movement, construction, or equipment staging

LLM Analysis

Each new image is analyzed by a vision LLM (Gemini) for aircraft count, vehicle concentration, runway status, and overall activity assessment.

Before/After comparison

When multiple observations exist for a site, the page shows side-by-side before/after images with dates. Hover over the “after” image to see spectral index values.

Limitations

• Cloud cover blocks optical imagery — SAR provides backup but is harder to interpret
• 10m resolution cannot distinguish individual vehicle types
• Seasonal changes (snow melt, vegetation) affect spectral indices
• LLM image analysis can hallucinate details