Estimation of PM2.5 Concentration Using Deep Bayesian Model Considering Spatial Multiscale

Abstract

Directly establishing the relationship between satellite data and PM2.5 concentration through deep learning methods for PM2.5 concentration estimation is an important means for estimating regional PM2.5 concentration. However, due to the lack of consideration of uncertainty in deep learning methods, methods based on deep learning have certain overfitting problems in the process of PM2.5 estimation. In response to this problem, this paper designs a deep Bayesian PM2.5 estimation model that takes into account multiple scales. The model uses a Bayesian neural network to describe key parameters a priori, provide regularization effects to the neural network, perform posterior inference through parameters, and take into account the characteristics of data uncertainty, which is used to alleviate the problem of model overfitting and to improve the generalization ability of the model. In addition, different-scale Moderate-Resolution Imaging Spectroradiometer (MODIS) satellite data and ERA5 reanalysis data were used as input to the model to strengthen the model’s perception of different-scale features of the atmosphere, as well as to further enhance the model’s PM2.5 estimation accuracy and generalization ability. Experiments with Anhui Province as the research area showed that the R2 of this method on the independent test set was 0.78, which was higher than that of the DNN, random forest, and BNN models that do not consider the impact of the surrounding environment; moreover, the RMSE was 19.45 μg·m−3, which was also lower than the three compared models. In the experiment of different seasons in 2019, compared with the other three models, the estimation accuracy was significantly reduced; however, the R2 of the model in this paper could still reach 0.66 or more. Thus, the model in this paper has a higher accuracy and better generalization ability.

Spatio-Temporal Characteristics of PM2.5 Concentrations in China Based on Multiple Sources of Data and LUR-GBM during 2016–2021

“…With the development of computer technology, machine learning (including deep learning) methods are increasingly used in the estimation of PM 2.5 concentrations due to their powerful non-linear modelling capabilities [16,17]. Such as support vector regression models [18,19], random forest models [20,21], artificial neural network models [22,23], Bayesian methods [24,25], generalised regression neural network models [26,27] and long and short-term memory networks [28], all of which have shown better performance than traditional statistical models in the estimation of PM 2.5 concentrations. In terms of the selection of influencing factors, these machine learning models used PM 2.5 information including adjacent temporal and spatial observations [29], land use information [30], vegetation index information [31], nitrogen dioxide (NO 2 ) concentration information [32], population density [33] and elevation [34], in addition to AOD and conventional meteorological observation parameters.…”