A study identifies regional and global hotspots of arsenic (As) contamination in topsoil using a deep learning model. Traditional methods for As detection are time-consuming and costly, but this research combines visible near-infrared (VNIR) spectra with a fully connected neural network (FCNN) to predict As content. The model achieved high accuracy (R² values of 0.688 and 0.692) and was used to estimate As levels at regional and global scales, identifying China, Brazil, and California as major hotspots. Other regions, such as Gabon, are at high risk but have been under-researched. The study highlights the need for targeted soil remediation and monitoring in these areas. The FCNN model offers a rapid, cost-effective alternative to traditional methods, enabling large-scale, non-destructive detection of As contamination. It was applied to US and global datasets, revealing high As content in various regions, including southern Thailand, China, Brazil, and Cuba. The model also identified populations at risk, with over 81 million people in the US potentially affected. The study emphasizes the importance of accurate As detection for environmental and public health protection. The FCNN model outperformed other machine learning approaches and provided insights into the spectral features influencing As prediction. The research underscores the need for a global soil spectral database and improved methods for assessing As contamination risks. The findings contribute to the development of strategies for sustainable soil management and environmental protection.A study identifies regional and global hotspots of arsenic (As) contamination in topsoil using a deep learning model. Traditional methods for As detection are time-consuming and costly, but this research combines visible near-infrared (VNIR) spectra with a fully connected neural network (FCNN) to predict As content. The model achieved high accuracy (R² values of 0.688 and 0.692) and was used to estimate As levels at regional and global scales, identifying China, Brazil, and California as major hotspots. Other regions, such as Gabon, are at high risk but have been under-researched. The study highlights the need for targeted soil remediation and monitoring in these areas. The FCNN model offers a rapid, cost-effective alternative to traditional methods, enabling large-scale, non-destructive detection of As contamination. It was applied to US and global datasets, revealing high As content in various regions, including southern Thailand, China, Brazil, and Cuba. The model also identified populations at risk, with over 81 million people in the US potentially affected. The study emphasizes the importance of accurate As detection for environmental and public health protection. The FCNN model outperformed other machine learning approaches and provided insights into the spectral features influencing As prediction. The research underscores the need for a global soil spectral database and improved methods for assessing As contamination risks. The findings contribute to the development of strategies for sustainable soil management and environmental protection.