Machine learning for predictive mapping of exceedance probabilities for potentially toxic elements in Czech farmland

For efficient decision-making and optimal land management trajectories, information on soil properties in relation to safety guidelines should be processed from point inventories to surface predictive maps. For large-scale predictive mapping, very few practical implementations have attempted to clarify how well indicator models can be built from large covariate sets combined with spatial proxies. This paper summarizes the performance of the weighted indicator-based random forest model which was used to predict exceedance probabilities for several potentially toxic elements (PTEs) in Czech farmland. The method was implemented for data mining in the Czech high-density monitoring data which had to be firstly regressed to achieve analytical harmony, and the reliability of the regression-based harmonisation was used as the input weights for the final model. The indicator-based models were trained for each PTE (As, Be, Cd, Co, Cr, Cu, Hg, Ni, Pb, V, and Zn) with two different sets of indicators, reflecting the two-tier nature of the Czech safety guidelines, which differentiate between soil textures of topsoil. The two separate predictive outputs are combined into a single probability map using a pragmatic meta-model of linear weights derived from a soil texture map generated by a compositional spatial model. Through validation with data splitting, the accuracy of the models showed relatively high predictive power for the probability distributions, but with pronounced differences between PTEs as the root mean square error in terms of exceedance probabilities ranged from 11 % (V) to 32 % (Cd and Cr) for independent validation. In addition, models based on high-resolution auxiliary variables allowed a meaningful and quantitative identification of the most important natural and anthropogenic drivers for areas with an increased rate of non-compliance with the protection thresholds for cultivated soils. Variable importance calculations showed the dominant influence of spatially explicit covariates (represented by geographical distances to quantile-based groups of points), but still significant contributions from other predictors. Among the natural factors, lithological information came to the fore, mainly due to continuous response variables such as mineral exploration density or geophysical ancillary variables (from remotely sensed gravimetry and radiometry). Among anthropogenic factors, particulate matter in the atmosphere was identified as the most important human-related pressure, followed by several land-use effects.