Machine learning based on the graph convolutional self-organizing map method increases the accuracy of pollution source identification: A case study of trace metal(loid)s in soils of Jiangmen City, south China

Abstract

Rapid economic development and industrialization may include environmentally harmful human activities that cause heavy-metal accumulation in soils, ultimately threatening the quality of the soil environment and human health. Therefore, accurate identification of pollution sources is an important weapon in efforts to control and prevent pollution. The self-organizing map (SOM) method is widely used in pollution source identification because of its capacity for visualization of high-dimensional data. The SOM ignores the graph structure relationship among chemical elements in soils; the SOM analysis of pollution sources has high uncertainty. Here, we propose a new analysis method, i.e., the graph convolutional self-organizing map (GCSOM), which uses a graph convolutional network (GCN) to extract the graph structure relationship among the chemical elements in soils, then performs data visualization using an SOM. We compared the performances of GCSOM and SOM, then assessed the pollution source characteristics of trace metal(loid)s (TMs, mostly heavy metals) in Jiangmen City using the GCSOM. Our experimental results showed that the GCSOM is superior to the SOM for identification of TM sources, while the TMs in the soil of Jiangmen originate from three main sources: agricultural activities (mainly in Taishan City, Jiangmen), traffic emissions (mainly in Xinhui and Pengjiang Districts), and industrial activities (mainly in Xinhui District). The risk assessment indicated that the risk of all TMs was within threshold.

Keywords: Graph convolutional network; Self-organizing map; Source analysis; Trace metal(loid)s.