Regional and global hotspots of arsenic contamination of topsoil identified by deep learning

Abstract

Abstract Topsoil arsenic (As) contamination threatens the ecological environment and human health. However, traditional methods for As identification rely on on-site sampling and chemical analysis, which are cumbersome, time-consuming, and costly. Here we developed a method combining visible near infrared spectra and deep learning to predict topsoil As content. We showed that the optimum fully connected neural network model had high robustness and generalization (R-Square values of 0.688 and 0.692 on the validation and testing sets). Using the model, the relative As content at regional and global scales were estimated and the human populations that might potentially be affected were determined. We found that China, Brazil, and California are topsoil As-contamination hotspots. Other areas, e.g., Gabon, although also at great risk, are rarely documented, making them potential hotspots. Our results provided guidance for regions that require more detailed detection or timely soil remediation and can assist in alleviating global topsoil-As contamination.