An Enhanced Geographically and Temporally Weighted Neural Network for Remote Sensing Estimation of Surface Ozone

Abstract

Surface ozone (O₃) pollution is a severe environmental problem that endangers human health. It is necessary to obtain high spatio-temporal resolution O₃ data to provide support for pollution monitoring and prevention. For this purpose, this study makes comprehensive use of remote sensing data, reanalysis data, and ground station observations, and develops an enhanced geographically and temporally weighted neural network (EGTWNN) model to acquire high spatial and temporal resolutions of O₃ data. The EGTWNN model is nested by two neural networks. The first neural network automatically learns the spatio-temporal proximity relationship to obtain spatio-temporal weights for the samples, and the spatio-temporal weights are then inputted into the second neural network to conduct weighted modeling of the relationship between O₃ and influencing variables. The contribution of the proposed model is that the first neural network replaces the traditional empirical weighting method, and represents the spatio-temporal proximity relationship more accurately to improve estimation accuracy. Results indicate that the cross-validation R² and RMSE of EGTWNN are 0.81 and 21.24 μg/m³, respectively, which are increased by 0.02 and decreased by ~1 μg/m³ relative to those of the traditional empirical weighting method based geographically and temporally weighted neural network model. The results also show that compared with the geographically and temporally weighted regression model, the proposed model achieves superior performance. In addition, the spatio-temporal weights obtained by the first neural network of EGTWNN are highly consistent with those obtained by the traditional empirical weighting method, indicating that the results of neural networks are highly interpretable.