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

Topsoil arsenic (As) contamination is a significant threat to both ecological health and human well-being. Traditional As detection methods are inefficient and expensive. This study presents a novel approach combining visible near-infrared (VNIR) spectroscopy and deep learning (specifically, a fully connected neural network, FCNN) to predict topsoil As content. The optimized FCNN model demonstrated high accuracy and generalizability (R-squared values of 0.688 and 0.692 on validation and testing sets). Using this model, regional and global topsoil As contamination was estimated, identifying China, Brazil, and California as hotspots. Other high-risk areas, such as Gabon, require further investigation. This method provides a rapid, cost-effective tool for identifying and addressing global topsoil As contamination.

Mengting Wu, Chongchong Qi, Sybil Derrible, Yosoon Choi, Andy Fourie, Yong Sik Ok