A framework for the predictive mapping of forest soil properties in mountain areas

Abstract

The modelling of forest ecosystems is a broad scientific field, encompassing species distribution, dynamic forest succession, growth and disturbance, and biogeochemical cycles. Soil information is frequently required for a holistic and spatially explicit modelling approach. Information on soil properties at sufficient resolution to be incorporated in spatially distributed models is rare however, in particular for mountain forest areas that are poorly accessible and where the required sampling effort is high. In order to bridge this gap, we aim to develop a concept for predicting spatially continuous forest soil properties in mountain areas and substantiate it by comparing different statistical approaches to digital soil mapping, namely Random Forests and Generalized Additive Models. Therefore, we used field descriptions of 1653 legacy soil profiles from a forest area of 5130 km2 in Tyrol and modelled a variety of soil properties that are essential for the characterisation of forest sites with respect to tree growth. A set of 23 spatially explicit predictor variables, grouped according to geological substrate information, topography, climate, biotic variables, and time–space, were included in the concept and tested. There was a special focus on the predictive relevance of the newly developed geological substrate information, which includes lithogenetic units, lithological composition, and the multilayering of sediments. As a physical soil property, it was possible to predict the plant-available water storage capacity with a significant degree of accuracy, with r2 = 0.49–0.56, while for the biologically driven state variable organic humus layer thickness, r2 = 0.22–0.27 was achieved. For describing soil reaction, Ellenberg’s indicator value of vegetation turned out to be a suitable proxy for soil pH value, and was modelled with a high degree of predictive power with r2 = 0.75–0.77. All model results were verified by different evaluation methods, ranging from cross-validation to comparison with independent datasets, tests for spatial autocorrelation, and considering ecological soundness in variable selection. We were able to show that in addition to variables derived from Digital Terrain Models and climate information, the new type of detailed geological substrate information is the most relevant predictor and is promising for digital soil mapping in mountain areas. This finding was independent of the applied statistical approaches and both Generalized Additive Models and Random Forests showed comparable accuracy and proved to be appropriate for these tasks. In addition, we were able to support the modelling results by interpreting the predictors’ mode of action with regard to the underlying processes controlling soil properties.