Soil & Water Res., X:X | DOI: 10.17221/128/2025-SWR

Balancing data quality in predictive geochemical mapping using machine learning: A Czech regional case study on topsoil nickelOriginal Paper

Jan Skála ORCID..., Daniel Žížala ORCID..., Robert Minařík ORCID...

¹ Research Institute for Soil and Water Conservation, Prague, Czech Republic

Machine learning makes geochemical mapping highly adaptable, as its data-driven nature allows predictions to evolve with new information. In this study, topsoil nickel (Ni) data were compiled from various sources, each with different sampling times and analytical methods. To effectively use such imbalanced data into spatial modelling, it was necessary to test how the data uncertainty propagated through the final maps. A comprehensive benchmark of the quantile random forest algorithm was conducted to identify conditions under which the model performs optimally. Predictive maps of topsoil Ni at a 20-metre resolution were subsequently generated and compared using a multi-faceted evaluation strategy. This approach assessed how model adjustments – particularly those addressing the uncertainty introduced by the regression-based conversion of legacy measurements – affected the performance. Extensive benchmarking revealed that while out-of-sample validation showed only modest improvements (e.g., root mean square error (RMSE) reduced from 12.6 to 11.2 mg/kg) when modifying training data, covariates, or algorithm parameters, the resulting prediction grids differed substantially. The analysis also demonstrated that output variability across model scenarios occurred at different spatial scales: weighting approaches had localised effects, whereas high variability in the input data propagated more broadly across the region.

Keywords: data uncertainty; prediction maps; topsoil geochemistry

Received: October 27, 2025; Accepted: March 10, 2026; Prepublished online: April 24, 2026