Causality-Driven Feature Selection for Calibrating Low-Cost Airborne Particulate Sensors Using Machine Learning.

Area of Science:

• Environmental Science
• Data Science
• Sensor Technology

Background:

• Escalating environmental challenges necessitate improved air quality monitoring.
• Low-cost sensor networks offer scalable solutions but require robust calibration.
• Traditional machine learning models struggle with interpretability and generalizability in environmental data.

Purpose of the Study:

• To develop a causal feature selection approach for enhanced machine learning calibration of low-cost sensors.
• To improve the interpretability and generalizability of environmental sensor data.
• To advance urban air quality monitoring through robust sensor network calibration.

Main Methods:

• Proposed a causal feature selection method using convergent cross mapping within a machine learning pipeline.
• Applied the approach to calibrate a low-cost optical particle counter (OPC-N3) for PM1 and PM2.5 measurements.
• Evaluated predictive performance and generalizability against traditional methods and SHAP value-based selection.

Main Results:

• Causally optimized models demonstrated improved predictive performance and generalizability for PM1 and PM2.5.
• Reduced mean squared error by 43.2% for PM1 and 33.2% for PM2.5 compared to models with all features.
• Outperformed SHAP value-based feature selection in reducing mean squared error for both PM1 and PM2.5 calibrations.

Conclusions:

• The causal feature selection approach enhances the robustness and interpretability of low-cost sensor network calibration.
• This method offers a significant advancement for urban air quality monitoring and understanding microenvironments.
• The approach has broad potential for other environmental monitoring applications, promoting interpretable environmental models.