This study developed a method to identify regional and global hotspots of arsenic (As) contamination in topsoil using deep learning. The method combines visible near-infrared spectra with a fully connected neural network (FCNN) model to predict As content. The optimal FCNN model achieved high robustness and generalization, with R² values of 0.688 and 0.692 on validation and testing sets, respectively. The model was applied to estimate As content at regional and global scales, identifying China, Brazil, and California as topsoil As-contamination hotspots. Other areas, such as Gabon, were also at risk but lacked documentation, making them potential hotspots. The study provides guidance for regions requiring detailed detection or timely soil remediation and contributes to alleviating global topsoil-As contamination. The FCNN model's performance was evaluated using various metrics, and its limitations and future improvements were discussed. The study highlights the importance of integrating diverse soil spectral data and expanding the global spectral library to enhance the model's accuracy and reliability.This study developed a method to identify regional and global hotspots of arsenic (As) contamination in topsoil using deep learning. The method combines visible near-infrared spectra with a fully connected neural network (FCNN) model to predict As content. The optimal FCNN model achieved high robustness and generalization, with R² values of 0.688 and 0.692 on validation and testing sets, respectively. The model was applied to estimate As content at regional and global scales, identifying China, Brazil, and California as topsoil As-contamination hotspots. Other areas, such as Gabon, were also at risk but lacked documentation, making them potential hotspots. The study provides guidance for regions requiring detailed detection or timely soil remediation and contributes to alleviating global topsoil-As contamination. The FCNN model's performance was evaluated using various metrics, and its limitations and future improvements were discussed. The study highlights the importance of integrating diverse soil spectral data and expanding the global spectral library to enhance the model's accuracy and reliability.