Retrieval of chromium and mercury concentrations in agricultural soils: Using spectral information, environmental covariates, or a fusion of both?

Abstract: The universal contamination of arable land with potentially toxic elements (PTEs) poses a severe threat to food security and jeopardizes worldwide efforts to meet the United Nations Sustainable Development Goals (SDGs). How to obtain information on PTEs in regional agricultural soils more reliably is a priority problem to be solved. Multispectral satellite remote sensing, with its advantages of high spatial and temporal resolution, broad coverage, and low cost, offers the potential to acquire distribution information of PTEs over large areas. However, owing to the complexity of soil environments and the insufficiency of spectral information, the mechanism for retrieving concentrations of soil PTEs via multispectral satellites is not yet clear, and the accuracy needs to be improved. In this study, we aimed to assess whether employing a fusion of spectral information and environmental covariates, results in more accurate predictions of PTEs, specifically chromium (Cr) and mercury (Hg), in croplands than does employing spectral information alone. Three machine learning algorithms—kernel-based support vector machine (SVM), neural network-based back propagation neural network (BPNN), and tree-based extreme gradient boosting (XGBoost)—were developed to retrieve soil Cr and Hg concentrations. The results showed that the fusion of spectral information and environmental covariates combined with the XGBoost model performed best in retrieving both Cr and Hg concentrations with coefficient of determination (R²) values of 0.73 and 0.74, respectively. Environmental covariates are important variables for determining Cr and Hg concentrations in agricultural soils, but the ability to retrieve these element concentrations by utilizing multispectral information alone is limited. High Cr concentrations occurred in central towns and southern hilly mountains. High Hg concentrations were detected in the northwestern region and southern hilly mountains. The potential of fusing multispectral data and environmental variables to precisely retrieve soil PTE concentrations can serve as a reference for agricultural information monitoring worldwide.

Keywords: Potentially toxic elements; Sentinel-2; Environmental covariates; Machine learning; Farmland