This project builds a spatio-temporal forecasting system that predicts urban noise in Barcelona by combining machine learning with geospatial analysis.
It integrates ArcGIS, GeoPandas, and NetworkX to engineer spatial context features (roads, parks, network centrality), allowing the model to “understand” the city’s physical structure before making predictions.
Objective: Forecast urban noise patterns and identify potential exceedances (> 65 dB) before they occur.
- Approach: Combine temporal forecasting models with geospatial context layers to improve prediction accuracy.
- Scale: Over 135 M sensor records processed and stored in a BigQuery + Cloud Storage data lake.
- Stack: Python (GeoPandas, Shapely, ArcGIS API, NetworkX, MLflow), Docker, GitHub Actions, Google Cloud Run.
Live Demo: Noise Forecasting App: sensor 496
The geospatial workflow forms the core of the project, enriching each sensor with urban context before feeding data into the forecasting model.
- Distance to main roads, green areas, and transport corridors using GeoPandas + Shapely
- Street-network betweenness centrality with NetworkX + OSMnx
- Local noise environment: neighborhood mean / variance within 150 m buffers
- Integration of all spatial features into a unified GeoDataFrame exported to the ML pipeline
- Automated spatial joins and geometry operations via ArcPy and ArcGIS API for Python
- Publication of geospatial layers and forecasted exceedances as interactive ArcGIS Online maps
- Visualization of predicted hot zones across Barcelona districts
- Detection of emerging noise clusters via Getis-Ord Gi* hotspot analysis and Moran’s I autocorrelation in ArcGIS
- Comparison of predicted vs. observed hotspots for model validation
ArcGIS Online Map: Noise Hotspot Analysis
ArcGIS Online Noise Hotspot Map
- Data Source: Noise Monitoring Network provided by OPEN DATA BCN, the open data portal of Barcelona City Council.
- Input: Noise sensor data (timestamp, location, dB levels)
- Feature engineering:
- Temporal features: hour, weekday, month, weekend flag
- Cyclical encoding (sin/cos)
- Lag features (1h, 24h), rolling statistics (3h, 24h)
- Data validation and cleaning to ensure valid input for modeling
- Models: Random Forest, Decision Trees (extensible design)
- Baselines: Persistence (last value), seasonal (24h lag)
- Backtesting:
- Expanding-window, one-step-ahead predictions
- Metrics: MAE, RMSE, relative improvement
- MLflow for experiment tracking, comparison, and production model selection
MLFlow Experiments saved locally for cost efficiency
- FastAPI: Serves real-time predictions and backtesting summaries
- Streamlit: Visual interface for model forecasts and error diagnostics
- Compatible with chat-based interfaces or voice-driven assistants
- Containerized with Docker
- CI/CD via GitHub Actions:
- Automatically builds and pushes Docker images to DockerHub
- Deploys to Google Cloud Run (previously deployed to AWS Fargate (ECS) but discontinued for cost-efficiency)
- Also adaptable to AWS/Azure infrastructure
- RMSE: 3.01 dB → ~70% lower error than the naive baseline (14.87 dB)
- MAE: 1.19 dB → below the human perception threshold (~3 dB)
- sMAPE: 1.75% | MASE: 0.80 → consistently better than persistence models
- Interval accuracy: ~93–95% coverage on 95% confidence → reliable, with room to fine-tune
- Stable and low-error forecasts (~1.2 dB) across time, even during noise pattern shifts.
- Outperforms baselines, enabling reliable detection of exceedances (e.g., >65 dB).
- Extend forecasting to all city sensors
- Integrate external factors like traffic and weather.