Neural Network Prediction of Pollutant Emissions from Open Burning of Crop Residues and Application to Air Quality Forecasts in Southern China

Open burning of crop residues is a strong seasonal source of air pollutants in many parts of China, but the large day-to-day variability of the associated emissions pose a great challenge for air quality forecasts. Here, we used back-propagation neural network (BPNN) ensembles to forecast the daily fire pixel counts in Southern China during the month of January. The BPNN ensembles were trained using a decade's worth (2003-2012) of daily assimilated surface meteorological data (including surface air temperature, relative humidity, surface pressure, and winds) and daily fire pixel observations from the Moderate Resolution Imaging Spectroradiometer (MODIS). We showed that the BPNN ensembles successfully forecasted the day-to-day variability of fire pixel counts over Southern China for January 2013 to 2015, with correlation coefficients of 0.6 to 0.8 against the MODIS observations. We then used the forecasted daily fire pixel counts to scale the climatological January biomass burning emissions from the Fire Inventory from NCAR (FINN) and applied the resulting daily biomass burning emissions to drive a regional air quality model. We compared air quality forecasts driven by our daily-variable emission inventory (called "BPNN_2014" case), as well as forecasts driven by the climatological January biomass burning emissions (called "FINN_MEAN" case). The normalized mean bias between the simulated PM2.5 concentrations from the two cases and observations were reduced from -8.2% to -0.3%. We showed that our daily-variable emission inventory led to significant improvements in the forecasts of PM2.5 and trace gas concentrations in Southern China.