Parsimonious Random-Forest-Based Land-Use Regression Model Using Particulate Matter Sensors in Berlin, Germany
Janani Venkatraman Jagatha1, Christoph Schneider1, Tobias Sauter1
1Geography Department, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany.
Sensors (Basel, Switzerland)
|July 13, 2024
Summary
Feature selection significantly improved machine learning models for predicting particulate matter (PM2.5) concentrations. Optimized models reduced errors and enhanced interpretability, highlighting key land-use predictors.
Area of Science:
- Environmental Science
- Data Science
- Atmospheric Science
Background:
- Machine learning (ML) models are crucial for particulate matter (PM2.5) prediction using air quality sensor data.
- The 'black-box' nature of ML models hinders understanding of prediction mechanisms, posing challenges in data quality and computational intensity.
- Feature selection methods are vital for improving the interpretability and efficiency of ML models in environmental applications.
Purpose of the Study:
- To employ feature selection techniques, specifically recursive feature elimination and global sensitivity analysis (GSA), to optimize a random forest (RF)-based land-use regression model for PM2.5 prediction in Berlin.
- To compare the performance of parsimonious RF models against a baseline model using various land-use predictors.
- To identify key land-use parameters influencing PM2.5 concentrations and enhance model interpretability.
Main Methods:
- Developed a baseline RF land-use regression model using 220 predictors for Berlin.
- Implemented five additional models: three using recursive feature elimination and two using Sobol-based GSA.
- Compared model performance using metrics like Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) against reference station data.
Main Results:
- Feature elimination reduced predictors from 220 to eight without performance loss.
- The GSA-based parsimonious model significantly outperformed the baseline, reducing MAE from 8.69 to 3.6 µg/m³ and RMSE from 9.86 to 4.23 µg/m³.
- The GSA_parsimonious model improved R² from 3% to 17% and predicted PM2.5 with MAE < 5 µg/m³ at 10 out of 12 reference locations.
Conclusions:
- Feature selection, particularly GSA, enhances ML model performance, interpretability, and efficiency for PM2.5 prediction.
- Population density, leaf area index, and traffic volume were identified as major PM2.5 predictors.
- While effective at local scales for identifying indicative land-use parameters, the GSA_parsimonious model's uncertainty limits regional-scale applicability.
