Geographically and temporally weighted neural networks for satellite-based mapping of ground-level PM2.5

Abstract

The integration of satellite-derived aerosol optical depth (AOD) and station-measured PM2.5 (particulate matter with an aerodynamic diameter of less than 2.5 μm) provides a promising approach for the monitoring of PM2.5. Previous models have generally only considered either the spatiotemporal heterogeneities of the AOD-PM2.5 relationship or the nonlinear relationship between AOD and PM2.5. In this paper, to simultaneously allow for the nonlinearity and spatiotemporal heterogeneities of the AOD-PM2.5 relationship, the geographically and temporally weighted neural network (GTWNN) model is proposed for the satellite-based estimation of ground-level PM2.5. The GTWNN model represents the nonlinear AOD-PM2.5 relationship via a generalized regression neural network, and is separately established for individual locations and times, to address the spatiotemporal heterogeneities of the AOD-PM2.5 relationship. Meanwhile, a spatiotemporal weighting scheme is incorporated in the GTWNN model to capture the local relations of samples for the training of the AOD-PM2.5 relationship. By the use of the Moderate Resolution Imaging Spectroradiometer (MODIS) AOD product, meteorological data, and MODIS normalized difference vegetation index (NDVI) data as input, the GTWNN model was verified using ground station PM2.5 measurements from China in 2015. The GTWNN model achieved sample-based cross-validation (CV) and site-based CV R2 values of 0.80 and 0.79, respectively, and it outperformed the geographically and temporally weighted regression model (CV R2: 0.75 and 0.73) and the daily geographically weighted regression model (CV R2: 0.72 and 0.72). The proposed model implements the combination of geographical law and a neural network, and will be of great use for remote sensing retrieval of environmental variables.