Abstract: Conventional assessment methods of soil heavy metal contamination predominantly depend on spatial interpolation analyses, conducted with data derived from a restricted number of sampling points, thereby often overlooking the influence exerted by environmental covariates such as geological backdrop, human activities, and geographical features on the heavy metal distribution. Predictions based solely on environmental covariates tend to fall short in adequately reflecting the spatial aggregation effects associated with heavy metal dispersion. As such, this paper introduces a novel approach that amalgamates spatial and environmental factors as covariates, employing three models: Random Forest (RF), Extreme Gradient Boosting (XGboost), and Deep Learning Neural Network (DNN), designed to predict the spatial distributions of heavy metals in Naozhou Island. The results underscore the substantial improvement in predictive performance achieved through the integration of spatial elements into the model. To test and enhance the robustness of the improved model, data splitting and external data validation techniques are utilized in this study, and the most optimal model for prediction is selected. Following the prediction results, correlation and clustering analyses were conducted on the heavy metal elements at the data feature level, and a LISA spatial clustering analysis at the spatial distribution level. The analyses reveal that geological factors predominantly dictate the dispersion of Cu, Ni, Cr, and Zn elements within the island, while human activities primarily govern the distribution of Pb, As, Hg. Moreover, the distribution of Cd element is ascertained to be influenced by a combination of human activities and geological background. 

Key words: Heavy metals, spatial heterogeneity, machine learning, accurate prediction