Exploring deep learning for air pollutant emission estimation

Abstract. The inaccuracy of anthropogenic emission inventories on a high-resolution scale due to insufficient basic data is one the major reasons for deviation between air quality model and observation results. A bottom-up approach, which typical inventory estimation method, requires lot human labor material resources, whereas top-down approach focuses individual pollutants that can be measured directly as well relying heavily traditional numerical modeling. Lately, deep neural network has achieved rapid development its high efficiency nonlinear expression ability. In this study, we proposed novel method dual relationship an pollution concentrations estimation. Specifically, utilized neural-network-based comprehensive chemical transport (NN-CTM) explore complex correlation pollution. We further updated based back-propagating gradient loss function measuring NN-CTM observations from surface monitors. first mimicked CTM with networks (NNs) relatively good representation CTM, similarity reaching 95 %. To reduce gap observations, NN suggests emissions NOx, NH3, SO2, volatile organic compounds (VOCs) primary PM2.5 changing, average, by −1.34 %, −2.65 −11.66 −19.19 % 3.51 respectively, in China 2015. Such ratios NOx are even higher (∼ 10 %) regions suffer large uncertainties original emissions, such Northwest China. improve performance make it closer observations. mean absolute error NO2, O3 reduced significantly (by about %–20 %), indicating feasibility terms improving both accuracy model.