SatQuery AI - SIH Prototype

An interactive web prototype for SatQuery AI: an agentic vision-language assistant for multimodal remote-sensing image analysis.

What this prototype implements

• Single-image natural-language analysis
• Remote-sensing caption/VQA-style workflow
• Text-guided region grounding prototype
• Bi-temporal before/after change analysis with a real pixel-difference evidence map
• Optical + SAR joint-analysis prototype with visual fusion overlay
• Input validation
• Agentic task routing
• Specialist tool registry
• Observable execution trace
• Confidence score
• Downloadable JSON analysis report
• GeoTIFF/TIFF/PNG/JPEG input acceptance at the API layer

Important scientific note

The build uses transparent image-processing baseline adapters for the demo. It does not falsely claim that these heuristics are a fine-tuned remote-sensing VLM. For the SIH evaluated build, replace the adapter functions in backend/analyzer.py with actual remote-sensing checkpoints/fine-tuned components and document the training/adaptation procedure.

The orchestration, API, UI, evidence system, and model-adapter interfaces can remain unchanged.

Run on Windows

1. Install Python 3.10+.
2. Open Command Prompt/PowerShell in this folder.
3. Create a virtual environment:

python -m venv .venv
.venv\Scripts\activate

4. Install dependencies:

pip install -r backend/requirements.txt

5. Start the server:

python backend/app.py

6. Open:

http://127.0.0.1:5000

Demo images

For the fastest demo, use PNG/JPEG satellite-style images. The interface accepts TIFF/GeoTIFF extensions as well. If you need full multispectral GeoTIFF band handling, add Rasterio/GDAL preprocessing in load_visual().

Recommended SIH demo sequence

Demo 1 — Single image

Upload one image and ask:

Describe the land-cover and major objects visible in this image.

Show:

• task classification
• RS captioning/VQA specialist
• answer
• confidence
• evidence

Demo 2 — Change analysis

Select Change Analysis, upload before/after images, ask:

What changed between these two dates, and where did the change occur?

Show:

• before
• after
• candidate change map
• execution trace

Demo 3 — Optical + SAR

Select Optical + SAR, upload optical + SAR images, ask:

Use the optical and SAR images together to identify built-up and water-covered regions.

Show:

• optical evidence
• SAR evidence
• joint overlay
• fusion integrator

How to upgrade the AI layer

The clean replacement points are:

• caption() → remote-sensing caption/VLM checkpoint
• VQA branch → remote-sensing VQA checkpoint
• grounding() → grounding/segmentation model
• change_analysis() → trained bi-temporal change detector + change captioner
• fusion_analysis() → optical-SAR fusion model

Keep the same JSON response contract:

{
  "answer": "...",
  "confidence": 0.88,
  "task": "...",
  "tools": ["..."],
  "evidence": [{"label": "...", "url": "..."}],
  "execution_trace": {}
}

This lets the UI remain stable while models are upgraded.

Suggested final architecture for judging

User → Input Validator → Agent Controller → Task Router → Specialist Model Registry → Model Execution → Evidence Generator → Result Integrator → Answer + Evidence + Confidence + Trace.

Presentation claim

"SatQuery AI is designed as an agentic orchestration layer rather than a single generic VLM. It validates the input configuration, classifies the user's intent, selects a specialist workflow, integrates textual and spatial evidence, and exposes an auditable execution trace."