Leveraging Unsupervised Data and Domain Adaptation for Deep Regression in Low-Cost Sensor Calibration.

Area of Science:

• Environmental Science
• Data Science
• Sensor Technology

Background:

• Air quality monitoring is crucial due to increasing environmental concerns.
• Low-cost sensors offer deployment advantages but lack the reliability of reference monitors.
• Deep learning presents a viable solution for calibrating low-cost sensors.

Purpose of the Study:

• To develop a novel semi-supervised domain adaptation method for calibrating low-cost air quality sensors.
• To address the challenges of covariate shift and label gap in sensor calibration.
• To enhance the reliability of low-cost air quality monitoring systems.

Main Methods:

• Framing sensor calibration as a semi-supervised domain adaptation problem.
• Utilizing histogram loss to mitigate covariate shift, replacing traditional Mean Squared Error (MSE) or Mean Absolute Error (MAE).
• Implementing sample weighting for adversarial entropy optimization to address the label gap.

Main Results:

• The proposed scheme significantly outperformed competitive semi-supervised and supervised domain adaptation baselines.
• Performance was validated using R-squared score and MAE metrics.
• Ablation studies confirmed the effectiveness of individual components within the proposed method.

Conclusions:

• The novel semi-supervised domain adaptation approach effectively calibrates low-cost air quality sensors.
• The method demonstrates superior performance compared to existing techniques, enhancing monitoring reliability.
• This research contributes to more accessible and accurate air quality assessment.