Modeling spatiotemporal hotspots and impact of cobalt contamination in European soils

Abstract

Soil cobalt contamination poses severe environmental and health risks. However, large-scale spatiotemporal variations in soil cobalt concentration remain poorly understood. Here, an attention-based deep learning model and four machine learning models were developed to generate cobalt concentration maps from soil spectra, covering European soils in 2009 and 2015. The dataset in 2009 comprised 18,675 samples, while the dataset in 2015 included 21,782 sample points. The deep learning model exhibited the optimal performance, with a mean squared error of 0.293, a mean absolute error of 0.146, and a coefficient of determination of 0.688, indicating robust predictive accuracy and model fit. Southern Europe, including Greece and Italy, had the highest soil cobalt concentrations. The validated model uncovered a significant increase in cobalt concentrations between 2009 and 2015, rising from a mean of 5.58 mg/kg to 10.49 mg/kg (P < 0.0001). Soil cobalt contamination (concentrations >20 mg/kg) was mainly found in Poland, Germany, and Italy in 2015, with Poland experiencing the largest increase. The model predicted that 1.5 million tons of crops produced and 12.3 million European people may have been potentially affected by cobalt contamination in 2015. This model can help identify regions requiring vigorous soil-remediation measures, while also raising awareness about reversing soil cobalt contamination.