---

Overview

SatQuery AI is an advanced multimodal remote-sensing and geospatial intelligence system developed for the ISRO AI Hackathon.

Rather than deploying a generic vision-language model, SatQuery implements an Autonomous Agentic Orchestration Pipeline. It interprets natural language intent, dynamically plans and coordinates specialized remote sensing analytical tools (Bi-Temporal Change Detection, Optical-SAR Fusion, Spatial Alignment, Coordinate Grounding), and queries a domain-adapted 7B Remote-Sensing VLM (RS-LLaVA-1.5-7B) fine-tuned on BigEarthNet and satellite VQA datasets.

Every response is anchored in a verifiable Evidence Graph, tying each natural language insight to exact pixel coordinates, Affine transforms, and deterministic tool outputs with zero hallucinations.

---

Key Capabilities

• Single-Image Remote Sensing VQA: Semantic scene interpretation, land-use categorization, object enumeration, and spatial grounding.
• Bi-Temporal Change Detection: Native support for GeoTIFF pairs (T1 and T2) with pixel-level differential analysis, bounding-box change localization, and semantic description.
• Cross-Modal Optical + SAR Fusion: Spatial co-registration and fusion of SAR backscatter intensity with Optical multispectral bands to analyze flooded regions, maritime structures, and obscured terrain.
• BigEarthNet-Adapted Neural Reasoning: Powered by BigData-KSU/RS-llava-v1.5-7b-LoRA loaded in 4-bit NF4 quantization for low-latency GPU inference (~5.5 GB VRAM footprint).
• Interactive Mission Control Dashboard: Glassmorphic web GUI with real-time pipeline tracking and evidence node inspection.
• Verifiable Evidence Graphs: Full compliance with the Absolute Integrity Rule—no fabricated inference or hallucinated geographic coordinates.

---

Benchmark Evaluation Results

SatQuery AI includes automated benchmark evaluation suites tested on real GPU hardware against standard Remote Sensing VQA datasets:

Raw benchmark telemetry and predictions are automatically exported to benchmark_results.json.

---

👨‍⚖️ Judges' Evaluation & Testing Guide

Judges can evaluate and test SatQuery AI using either Google Colab (Free GPU) or Local CLI / Docker.

---

Method 1: Instant Evaluation on Google Colab (Recommended)

To run the complete system with the 7B Neural VLM on a free T4 GPU in Google Colab:

Step 1: Clone and Set Up Environment

!git clone https://github.com/ainaanraza/SatQuery.git /content/GeoChat
%cd /content/GeoChat
!pip install fastapi uvicorn pydantic rasterio transformers accelerate bitsandbytes peft

Step 2: Run Dataset Benchmarks

You can evaluate the model against VRSBench or any remote-sensing dataset split:

# Run 20-sample benchmark evaluation on VRSBench
!python -m satquery.evaluation.benchmark --dataset /content/data/VRSBench --provider hf_llava --samples 20

Step 3: Launch Live API & Tunnel for Web GUI

Run this Python snippet to start the server in the background and generate a live tunnel:

import subprocess, time, urllib.request, re, os

os.environ["SATQUERY_MODEL_PROVIDER"] = "hf_llava"
os.environ["SATQUERY_ENV"] = "production"

!pkill -9 -f uvicorn
!pkill -9 -f cloudflared

if not os.path.exists("/content/cloudflared-linux-amd64"):
    !wget -q -O /content/cloudflared-linux-amd64 https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64
    !chmod +x /content/cloudflared-linux-amd64

subprocess.Popen(["python", "-m", "uvicorn", "satquery.api.app:app", "--host", "0.0.0.0", "--port", "8000"], cwd="/content/GeoChat")
time.sleep(3)
subprocess.Popen(["/content/cloudflared-linux-amd64", "tunnel", "--url", "http://127.0.0.1:8000", "--metrics", "127.0.0.1:8099"])

for _ in range(15):
    time.sleep(1)
    try:
        metrics = urllib.request.urlopen("http://127.0.0.1:8099/metrics", timeout=2).read().decode()
        match = re.search(r'userHostname="(https://[a-zA-Z0-9\-]+\.trycloudflare\.com)"', metrics)
        if match:
            print(f"\n🚀 LIVE API URL: {match.group(1)}")
            break
    except Exception:
        pass

1. Copy the printed https://...trycloudflare.com URL.
2. In satquery/frontend/app.js, set const API_BASE_URL = '<YOUR_URL>'; (Line 1).
3. Open satquery/frontend/index.html in your browser to test the interactive mission control dashboard.

---

Method 2: Testing on Different Public Datasets

SatQuery AI provides native recursive evaluation loaders for major remote sensing benchmark datasets:

1. Testing on VRSBench:

# Clone VRSBench test suite
git clone https://github.com/lx709/VRSBench.git /content/data/VRSBench

# Run benchmark evaluation
python -m satquery.evaluation.benchmark --dataset /content/data/VRSBench --provider hf_llava --samples 50

2. Testing on RSVQA:

python -m satquery.evaluation.benchmark --dataset /path/to/RSVQA --provider hf_llava --samples 50

3. Testing on CDVQA (Change Detection VQA):

python -m satquery.evaluation.benchmark --dataset /path/to/CDVQA --provider hf_llava --samples 50

4. Testing on Custom GeoTIFF Satellite Images:

Place your own .tif / .tiff satellite files into the directory and evaluate:

python -m satquery.agent_cli --query "Describe the agricultural density, water bodies, and roads" --image demo_satellite.tif

---

Method 3: Local Command-Line Interface (CLI)

You can run agent queries directly from your terminal:

# Single-image remote sensing query
python -m satquery.agent_cli --query "Identify all maritime structures and docks" --image test1.tif

# Bi-temporal change detection query
python -m satquery.agent_cli --query "Detect new construction between T1 and T2" --image-before demo_satellite.tif --image-after test1.tif

# Generate synthetic test rasters
python -m satquery.create_demo_raster

---

Modular System Architecture Breakdown

SatQuery AI Architecture:
├── Geospatial Core:     Memory-Safe Raster Ingestion (Rasterio), Multi-Band Normalization, CRS
├── Agentic Core:        Intent Parser, Dynamic Task Planner, Capability-Based Executor
├── Analytical Tools:    Optical/SAR Fusion, Bi-Temporal Change Detection, TemporalSeries Engine
├── Intelligence Layer:  HuggingFace Provider Registry, 4-Bit RS-LLaVA VLM (BigEarthNet)
├── Verification Layer:  Verifiable Coordinate Grounding, EvidenceGraph, Zero-Hallucination Guard
└── Interface Layer:     FastAPI Async Job API, Mission Control Glassmorphic Web Dashboard

---

Repository Structure

├── satquery/
│   ├── agent/                 # Autonomous Agent (Planner, IntentParser, Executor)
│   ├── api/                   # FastAPI REST server endpoints
│   ├── evaluation/            # BenchmarkEngine, VRSBench/RSVQA evaluation suite
│   ├── evidence/              # Evidence Graph data structures & verification
│   ├── frontend/              # Mission Control Dashboard (HTML, CSS, JS)
│   ├── inputs/                # Memory-safe GeoTIFF raster loaders (Rasterio)
│   ├── models/                # Provider abstraction & RS-LLaVA 4-bit integration
│   ├── tools/                 # Change detection, SAR fusion, preview tools
│   └── create_demo_raster.py  # Multi-band GeoTIFF test generator
├── docs/                      # Phase audit and architecture reports
├── benchmark_results.json     # Official VRSBench benchmark output artifact
└── README.md

---

Built for the ISRO Remote Sensing.