Alt title: Chicago Urban Mobility Intelligence: Integrating Satellite Imagery, Taxi Trajectory Data, and Machine Learning for Urban Demand Analysis
A production-ready geospatial analytics platform that combines satellite imagery, machine learning, residual diagnostics, and an LLM-powered spatial analyst for urban mobility forecasting in Chicago.
The system integrates multi-source spatial data, static and temporal XGBoost models, interactive mapping, and a tool-using conversational agent into a single Streamlit application.
Access the app here 👉: https://bit.ly/4wNJQx1
How it works: https://youtu.be/VqZ9-ZvSHDY
Final platform view
Dashboard Preview
Chicago Urban Mobility Intelligence is an end-to-end decision-support platform for exploring and forecasting taxi demand across Chicago’s 77 community areas.
It brings together:
- Sentinel-2 satellite-derived vehicle density proxies
- 14M+ Chicago taxi trip records
- OpenStreetMap road network and business/POI data
- Population density (WorldPop)
- Static (community-area) and temporal (hourly) XGBoost models
- Residual analysis for hidden demand hotspots
- An LLM-powered AI Mobility Analyst with tool calling
The platform is designed for both research insight and operational exploration (maps, forecasts, and conversational spatial analysis).
The main Dashboard tab provides:
- Interactive Folium maps with multiple layers (Total Taxi Trips, Trips per 1,000 People, Satellite Proxy, Residuals)
- Custom basemap options (CartoDB Positron, Satellite, OpenStreetMap)
- Quick Temporal Forecast panel (select community area, hour, and weekend flag)
- Automatic map highlight and zoom to the predicted community area
- Live prediction metric display
Users can explore spatial demand patterns and generate hourly forecasts in one interface.
The AI Mobility Analyst tab pairs an interactive map with a tool-using LLM agent (Groq / Llama 3.3).
The agent can:
- Predict hourly taxi demand for any community area
- Retrieve highest-demand neighborhoods
- Report spatial model feature importance
- Identify hidden hotspots (positive residuals)
- Return detailed statistics for a selected area
- Automatically update map layer and focus based on the query
Responses are analytical (not raw tool dumps) and oriented toward transportation planning insight.
- Satellite: Sentinel-2 MSI (10 m bands B02, B03, B04, B08) for texture / vehicle density proxies
- Mobility: Chicago Taxi Trips (~14M records) as ground truth
- Roads: OpenStreetMap network extract
- Population: WorldPop 2020 gridded estimates
- POI / Business: Chicago Business Licenses (hotels, bars, restaurants, etc.)
- Spatial feature engineering (POI counts, airport flag, distance to Loop, satellite proxy, population density)
- Static XGBoost model → annual demand per community area
- Residual analysis → hidden hotspots and over-prediction zones
- Temporal aggregation (hour × day type × community area)
- Temporal XGBoost model (native DMatrix) → hourly demand forecasts
- Interactive Streamlit interface + LLM agent with tool calling
| Model | Metric | Value |
|---|---|---|
| Static (community) | R² | 0.666 |
| Static | MAE | 50,806 trips |
| Temporal (hourly) | R² | 0.954 |
| Temporal | MAE | 294 trips |
- num_hotels — 0.487
- is_airport — 0.217
- dist_to_loop_km — 0.082
- num_bars — 0.062
- num_restaurants — 0.055
Main findings
- Taxi demand is strongly destination-driven (hotels and airports dominate).
- Sentinel-2 10 m texture proxy adds limited predictive value once POI and location features are included.
- O’Hare and Near North Side remain significant under-predicted (hidden) hotspots.
- Several far South Side areas are systematically over-predicted.