Deep-BCSI: A deep learning-based framework for bias correction and spatial imputation of PM2.5 concentrations in South Korea

Abstract

In this study, we introduce a deep learning-based framework, Deep-BCSI, which leverages Convolutional Neural Networks (CNN) for bias correction and Partial Convolutional Neural Networks (PConv) for spatial imputation. It is designed to enhance the accuracy of PM2.5 concentration forecasts over South Korea, at both station and grid levels, up to three days in advance. The framework utilizes 72-h simulations of 10 variables from the Community Multiscale Air Quality (CMAQ) model, 31 variables from the Weather Research and Forecasting (WRF) model, and 6 variables from the previous day's ground-based in-situ observations. The CNN and PConv models' training time spans from 2016 to 2019. The Deep-BCSI framework was evaluated in 2021. From Day 1 to Day 3, the CNN model demonstrated significant efficiency in bias correction, yielding higher Index of Agreement (IOA) values (0.71–0.80) compared to the CMAQ's (0.65–0.68) across 402 stations in South Korea. In metropolitan areas such as Seoul, Busan, Incheon, and Daegu, the Root Mean Squared Error (RMSE) was reduced by 25% - 41%. Post-bias correction, spatial imputation using the PConv model provided accurate grid-based forecasts of PM2.5 concentrations, particularly in the northwestern and southeastern regions of South Korea. From Day 1 to Day 3, a 10-fold spatial cross-validation reveals that the PConv model consistently yielded higher IOA values (0.71–0.79) compared to the CMAQ's (0.66–0.69). An analysis using Shapley Additive Explanations (SHAP) offered insights into the CNN model's prediction-making process, confirming it as scientifically valid and closely aligned with essential atmospheric chemistry and meteorological phenomena related to PM2.5 pollution. The Deep-BCSI presents an efficient framework for generating accurate PM2.5 concentration forecasts, particularly for urban regions. It has potential applications in operational facilities, planning, and policymaking to mitigate the hazards posed by PM2.5 pollution. © 2023