Band-Sensitive Calibration of Low-Cost PM2.5 Sensors by LSTM Model with Dynamically Weighted Loss Function

Abstract

Particulate matter has become one of the major issues in environmental sustainability, and its accurate measurement has grown in importance recently. Low-cost sensors (LCS) have been widely used to measure particulate concentration, but concerns about their accuracy remain. Previous research has shown that LCS data can be successfully calibrated using various machine learning algorithms. In this study, for better calibration, dynamic weight was introduced to the loss function of the LSTM model to amplify the loss, especially in a specific band. Our results showed that the dynamically weighted loss function resulted in better calibration in the specific band, where the model accepts the loss more sensitively than outside of the band. It was also confirmed that the dynamically weighted loss function can improve the calibration of the LSTM model in terms of both overall performance and local performance in bands. In a test case, the overall calibration performance was improved by about 12.57%, from 3.50 to 3.06, in terms of RMSE. The local calibration performance in the band improved from 4.25 to 3.77. Such improvements were achieved by varying coefficients of the dynamic weight. The results from different bands also indicated that having more data in a band will guarantee better improvement.

Enhancing Global Estimation of Fine Particulate Matter Concentrations by Including Geophysical a Priori Information in Deep Learning

“…The parameter y i is the label data in pixel i , i.e., the bias of the geophysical a priori PM 2.5 versus ground monitor PM 2.5 . N is the total number of training datasets input to the model in one training step.Few previous studies have customized the loss function for deep learning methods in PM 2.5 estimation. − In this study, we identified shortcomings associated with employing the MSE loss function for training a global CNN PM 2.5 model. These deficiencies encompass disregard for regions characterized by low PM 2.5 concentrations and the production of unreasonable estimations in areas with limited monitor coverage.…”