CNN-based model for the spatial imputation (CMSI version 1.0) of in-situ ozone and PM2.5 measurements

Abstract

This research presents a unique application of the convolutional neural network (CNN) for imputing data of hourly ozone concentrations between surface air quality stations. The study's objective is to develop a model that can prepare a 2D gridded map from the available observation from ground measurement. We train the CNN-based Model for Spatial Imputation (CMSI) with the Community Multiscale Air Quality (CMAQ) model-simulated hourly ozone concentrations. The modeled ozone with the CMAQ at a 27 km horizontal resolution acted as a proxy for the spatially complete and gridded ozone dataset. We also evaluated the CMSI models spatially for in-situ measurement using a 10-fold evaluation, and we found a yearly index of agreement of 0.91, a mean bias of 3.06 ppb, and a root mean square error of 10.89 ppb (an average of combined 10 sets of evaluations). In addition, we also trained two different CMSI models as case studies; i) for PM2.5 concentrations to assess its robustness for other atmospheric constituents, and ii) for regridded fine resolution (0.1⁰ × 0.1⁰) ozone concentrations. The performance of the CMSI model for the imputation of in-situ measurements was independent of the proxy model used for the specific resolution, however, using a proxy model with better physics and chemistry may improve the performance of the CMSI model. The system demonstrates the capability to accurately impute surface ozone concentrations within grids for which observation data are not available. The spatially gridded dataset generated by the CMSI model can further be used in conjunction with a numerical model for estimation and forecasting. The CMSI model can also be used for creating fine resolution gridded datasets based on in-situ measurement.