Deep Learning Prediction of Polycyclic Aromatic Hydrocarbons in the High Arctic.

Given the lack of understanding of the complex physiochemical and environmental processes of persistent organic pollutants (POPs) in the Arctic and around the globe, atmospheric models often yield large errors in the predicted atmospheric concentrations of POPs. Here, we developed a recurrent neural network (RNN) method based on nonparametric deep learning algorithms. The RNN model was implemented to predict monthly air concentrations of polycyclic aromatic hydrocarbons (PAHs) at the high Arctic monitoring station Alert. To train the RNN system, we used MODIS satellite remotely sensed forest fire data, air emissions, meteorological data, sea ice cover area, and sampled PAH concentration data from 1996 to 2012. The system was applied to forecast monthly PAH concentrations from 2012 to 2014 at the Alert station. The results were compared with monitored PAHs and an atmospheric transport model (CanMETOP) for POPs. We show that the RNN significantly improved PHE and BaP predictions from 2012 to 2014 by 62.5 and 91.1%, respectively, compared to CanMETOP predictions. The sensitivity analysis using the Shapley value reveals that air emissions determined the magnitude of PAH levels in the high Arctic, whereas forest fires played a significant role in the changes in PAH concentrations in the high Arctic, followed by air temperature and meridional wind fields.