Enhancing Machine Learning with Low-Cost P M2.5 Air Quality Sensor Calibration using Image Processing

Abstract

Low-cost particulate matter sensors, due to their increased mobility compared to reference monitors, are transforming air quality monitoring. Calibrating these sensors requires training data from reference monitors, which is traditionally done through conventional procedures or by using machine learning techniques. The latter outperforms traditional methods, but still requires deployment of a reference monitor and significant amounts of training data from the target sensor. In this study, we present a cutting-edge machine learning-based transfer learning technique for rapid sensor calibration with Co-deployment with reference monitors is kept to a minimum. This approach integrates data from a small number of sensors, including the target sensor, reducing the dependence on a reference monitor. Our studies reveal that In recent research, a transfer learning method using a meta-agnostic model has been proposed, and the results proved to be much more effective than the previous method. In trials, calibration errors were successfully reduced by up to 32\% and 15\% compared to the best raw and baseline observations. This shows the great potential of transfer learning methods to increase the effectiveness of learning in the long term. These results highlight the potential of this innovative transfer learning technique for rapidly and accurately calibrating low-cost particulate matter sensors using machine learning.