|
Surface ozone can trigger many health problems for human (e.g. coughing, bronchitis, emphysema, and asthma), especially for children and the elderly. It also has harmful effects on plants (e.g. chlorosis, necrosis, and yield reduction). The United State (U.S.) Environmental Protection Agency (EPA) has been monitoring surface ozone concentrations across the U.S. since 1980s. However, their stations are sparsely distributed and mainly in urban areas. Evaluation of surface ozone effects at any given locations in the U.S. requires spatial interpolation of ozone observations. In this study, we implemented two traditional spatial interpolation methods (i.e. triangulation-based linear interpolation and geostatistics-based method). One limitation of these two methods is their reliance on single-scene observations in constructing the spatial relationship, which is prone to influence of noisy observations and has large uncertainty. Deep learning, on the other hand, is capable of simulating common patterns (including complex spatial patterns) from a large amount of training samples. Therefore, we also implemented three deep learning algorithms for the spatial interpolation problem: mixture model network (MoNet), Convolutional Neural Network for Graphs (ChebNet), and Recurrent Neural Network (RNN). The training and validation data of this study are the 2016 EPA hourly surface ozone observations within ±3-degree box centered at the Billings, Oklahoma station (USDA UV-B Monitoring and Research Program). The results showed that among the five methods, RNN and MoNet outperformed the two traditional spatial interpolation methods and RNN has the lowest validation error (mean absolute error: 2.82 ppb; standard deviation: 2.76 ppb). Finally, we used the integrated gradients method to analyze the attribution of RNN inputs on the surface ozone prediction. The results showed that surface ozone observation is the most important input feature followed by distance and absolute locations (i.e. elevations, longitudes, and latitudes).
|
