Variability across scales - exploring methods for predicting soil properties from multiple sources

To assess future groundwater recharge rates in Austria under climate change conditions, detailed spatial soil information is required.  Different data sources such as global soil maps (SoilGrids), regional soil maps of arable land (eBOD) and local soil profiles are available. However, they differ in scale and degree of data aggregation, as well as in spatial coverage.

Soil properties are characterized by a high spatial variability at all scales and it is well known that averaging will cause biases in the statistical relationships between different soil data sets, and between soil and landscape physio-geographical properties. Aiming for a best quality Austrian-wide soil map synthesizing all information, scale dependent multi-level relations between soil data bases are examined following two approaches:

Firstly, a linear relationship of soil variables at different scales is assumed. The statistical and geostatistical characteristics are analyzed at different aggregation levels to explore the scale-related behavior of our data. Secondly, machine learning algorithms (random forests and boosting methods) are applied to predict soil characteristics of existing regional soil maps by using all other available data sources as input features. Additional locally available variables such as elevation, slope, aspect, vegetation and climate data are evaluated for significance when predicting the missing soil information.  

In summary, this study analyzes the statistical behavior and patterns of variability of soil properties at different scales and derives a modelling approach that can be used to predict regional soil properties from data sources spanning a range of different scales.