Spatio-temporal Pattern Of Wet Nitrogen Deposition In Guangdong Province Based On Deep Learning Method

Regional atmospheric nitrogen deposition simulation and prediction is a research hotspot in the field of atmospheric pollution research.The rapid increase in nitrogen deposition rates on a global scale poses a serious threat to ecosystems,causing ecological problems including soil acidification,eutrophication of water bodies,low biodiversity,and so on.The most common method of nitrogen deposition simulation is technical analysis based on observation data and numerical models.However,the observation data are limited when it comes to analysis of spatial patterns due to the limited number of observation sites.Deposition simulation using numerical models,especially wet deposition simulation,has been a difficult bottleneck to break through in the field of numerical simulation of atmospheric pollution due to several problems,such as the large uncertainty of precipitation simulation,and the high dependence on the accuracy of emission inventory.Machine learning（deep learning）method can be an effective complementary tool when numerical model simulation results are not ideal.In this paper,the application of machine learning（deep learning）in wet deposition simulation was explored.Besides,the spatial and temporal patterns of wet deposition fluxes from 2010 to 2017 in Guangdong province based on machine learning（deep learning）simulation results was systematically analyzed.In addition,the influence of weather conditions in wet deposition in Guangdong Province are also quantified using the classification method named“Lamb-Jenkinson weather types”.The main conclusions obtained from the study are as follows.（1）The machine learning（deep learning）models selected in this paper were significantly better than the deposition numerical model WRF-EMEP（Weather Research and Forecasting-European Monitoring and Evaluation Programme）.Among the selected machine learning（deep learning）models,the Convolutional Neural Network（CNN）was better than the other three types of machine learning（deep learning）models and showed the best spatial generalization ability.The machine learning（deep learning）models have good application prospects in the simulation of nitrogen wet deposition at the regional scale.（2）The simulation results of the optimal deposition simulation model,CNN,showed the large difference between spatial and temporal patterns of different nitrogen deposition species in Guangdong Province.The wet deposition of oxidized nitrogen（NO₃^-）was mainly distributed in the northern and central Pearl River Delta,while the deposition of reduced nitrogen（NH₄⁺）was more distributed in the central and northern Guangdong province.The high flux of reduced nitrogen deposition in northern Guangdong needs to be a concern.In terms of interannual and seasonal characteristics,the highest deposition year was 2016 during the period,and the high deposition seasons for two deposition species were spring and summer,with relatively less deposition in autumn and winter.（3）The main weather types affecting wet deposition in Guangdong province were E,SE,C,S,and CE,with the total contribution of 70.8%to the deposition flux.E type was the main weather type affecting wet deposition,accounting for 27.0%of the total flux.Besides,seasonal differences were obvious in the influence of weather type on wet deposition.The weather types corresponding to the deposition fluxes in spring and summer were mainly C type and the corresponding hybrid types,S and the corresponding hybrid types,while autumn and winter were mainly controlled by E type.