Soil information on a regional scale: Two machine learning based approaches for predicting saturated hydraulic conductivity

Saturated hydraulic conductivity (Ksat) and other soil (hydraulic) properties are fundamental for applications that depend on modeling hydrological processes, such as the quantification of future groundwater recharge rates. Yet, for most areas in the world, local soil information is lacking. Additionally, access to local soil surveys is often restricted or costly. Available global and regional digital soil mapping (DSM) products differ in scale and degree of data aggregation, as well as in spatial coverage. Ksat – and soil properties in general – are also characterized by a high spatial variability at all scales. Most often, there is no single data product available that covers the whole study area and still displays the variability of local soil observations. Thus, it is often a challenge to combine and predict soil data from different sources and resolutions while preserving the characteristically high spatial variability of soil properties. This study develops and compares two approaches for producing spatially distributed Ksat maps. First, an indirect approach based on two machine learning (ML) models – eXtreme Gradient Boosting (XGBoost) and feed-forward neural network (FNN) – that are trained with available local soil data sources and environmental raster datasets to predict the soil parameters sand, silt, clay, and organic matter content. Ksat is then determined by applying existing pedotransfer-functions (PTFs) on these regionalized soil parameters. Second, a direct approach in which ML models are directly trained with available soil hydraulic datasets to predict Ksat. Both approaches are applied to predict Ksat for Austria. While the resulting soil property maps of the indirect approach are able to largely reproduce the original data variability, the prediction of Ksat includes high levels of uncertainties and the predicted vertical distribution of Ksat is not plausible. The spatial distribution of Ksat in the direct approach resembles available global Ksat maps. In the existing global Ksat maps as well as in the results of the direct approach the small-scale variability of Ksat is reduced. In both approaches XGBoost outperforms FNN. The derived soil property maps help to reduce current gaps in soil data availability for Austria, but also highlight the need for additional Ksat field data acquisition.