🛰️ NAIP Chat — Conversational Earth Observation with Qwen3-VL

A Streamlit app that lets you fetch free 0.6-meter aerial imagery of any location in the continental United States and have a real-time conversation about what's in it — powered by the Qwen3-VL 235B vision-language model.

🔗 Live Demo: https://naipchat.streamlit.app

---

What It Does

1. Enter any latitude/longitude (defaults to Catonsville, MD)
2. Fetch a NAIP aerial tile from Microsoft Planetary Computer — free, no account needed
3. Ask the AI anything about the image in natural language
4. Get expert-level geospatial analysis streamed back in real time

The model is prompted as a remote sensing scientist, so responses focus on land cover, infrastructure, vegetation patterns, impervious surfaces, water features, and real-world geographic context — not generic image captions.

---

Demo Questions to Try

• "What is the dominant land cover type in this image?"
• "Describe the road network and connectivity visible here."
• "How much of this tile appears to be impervious surface?"
• "Are there any signs of vegetation stress near water bodies?"
• "What kind of neighborhood is this — urban, suburban, or rural fringe?"
• "Describe any visible agricultural field patterns."

---

Stack

Component	Tool
Aerial imagery	NAIP via Microsoft Planetary Computer
Vision-language model	Qwen3-VL 235B via Ollama Cloud API
Image I/O	Rasterio, Pillow
UI	Streamlit
API client	OpenAI-compatible (openai Python SDK)

---

Why NAIP?

NAIP (National Agriculture Imagery Program) is collected by the USDA at 0.6-meter resolution across the continental US every 2–3 years. At that resolution, individual trees, building footprints, parked cars, and vegetation edges are clearly visible — making it far more useful for parcel-level analysis than Sentinel-2 (10m) or Landsat (30m).

---

Local Setup

1. Clone the repo

git clone https://github.com/yourusername/naipchat.git
cd naipchat

2. Install dependencies

pip install -r requirements.txt

Note: On Windows, install rasterio via conda: conda install -c conda-forge rasterio

3. Configure secrets

Create .streamlit/secrets.toml:

OLLAMA_API_KEY = "your_ollama_api_key"
OLLAMA_HOST    = "https://ollama.com"
OLLAMA_MODEL   = "qwen3-vl:235b-cloud"

Get your Ollama API key at ollama.com.

4. Run

streamlit run app.py

---

Deploying to Streamlit Cloud

1. Push repo to GitHub
2. Connect at share.streamlit.io
3. Add the three secrets above under App Settings → Secrets
4. Add a packages.txt file to the repo root for rasterio's GDAL dependency:

libgdal-dev

---

Requirements

streamlit>=1.35.0
openai>=1.30.0
pystac-client>=0.7.0
planetary-computer>=1.0.0
rasterio>=1.3.0
Pillow>=10.0.0
numpy>=1.26.0

---

Features

• 📍 Configurable AOI — any lat/lon in the continental US, adjustable buffer radius
• 🖼️ Live tile preview — NAIP image displayed alongside the chat
• 💬 Persistent conversation — multi-turn chat with full context retained
• 🧠 Editable system prompt — swap analyst personas in the sidebar without touching code
• ⚡ Streaming responses — answers appear token by token in real time
• 🆓 No imagery account needed — Planetary Computer NAIP access is fully open

---

Limitations

• NAIP coverage is US only and updated every 2–3 years per state — recently changed areas may not reflect current conditions
• The model is a general-purpose VLM, not fine-tuned on remote sensing benchmarks — responses are best suited for scene interpretation and Q&A, not precise pixel-level classification
• Large buffer values may produce very large image tiles and slow API responses

---

License

MIT

---

Author

Built by R K — geospatial data scientist and Earth observation researcher