Drivers distinguishing of PAHs heterogeneity in surface soil of China using deep learning coupled with geo-statistical approach

Although numerous studies have reported the influencing factors of polycyclic aromatic hydrocarbons (PAHs) in surface soil from source, process or soil perspectives, the mechanism of PAHs heterogeneity in surface soil are still not well understood. In this study, the effects of 16 PAHs in surface soil of China sampled between 2003 and 2020 with their 17 "source-process-sink" factors at 1 km resolution (N = 660)) were explored using deep learning (eXtreme Gradient Boosting) to mine key information from complex dataset under the optimized parameters (i. e., learning rate = 0.05, maximum depth = 5, sub-sample = 0.8). It was observed that top five factors of 16 PAH had the largest cumulative contribution (i.e., from 84.8% to 98.1%) on their soil concentrations. PAH emission was the predominant driver, and its effect on soil PAH increases with increasing logKow. Soil was the second driver, in which clay can promote the partition of PAHs with low or middle logKow. However, sand can accumulate those congeners with high logKow. Moreover, the deep learning plus geo-statistical models (with low deviation for testing dataset (N = 283)) were capable of predicting soil PAH concentrations using their drivers with high accuracy. This study improved the understanding of the environmental fate and spatial variability of soil PAHs, as well as provided a novel technique (i.e., deep learning coupled with geo-statistics) for accurate prediction of soil pollutants.