Mapping soils in temperate humid mountains: Random forests highlights optimal covariates in Northern Italy.

As protection and conservation of habitats gain importance in the environmental policies of many countries, knowledge of the spatial distribution of soil types is increasingly needed to manage natural and semi-natural lands. Mountainous areas have high environmental variability, and this complexity can make the study of such landscapes particularly challenging. Recently, digital soil mapping (DSM) has gained popularity with the increasing availability of data layers that, together with soil observations and machine learning techniques such as random forests, allow the production of accurate soil maps. In this study we used random forests with recursive feature elimination to map the distribution of soil types in two protected natural parks in the mountains of northern Italy. We used 108 soil pedon observations and an optimal subset of five covariates (elevation, diffuse insolation, NDVI, slope, and distance from lakes) to make several predictions of soil functional group distribution. The final soil map created using the modes of the predictions was accurate (Kappa index = 0.82) and could be confidently used by land managers, while identifying the leading pedogenetic factors in this landscape. Furthermore, running several predictions allows the mapping of soil transition zones and complex areas which might require more careful management. Knowing which covariates are more influential in this landscape allows the optimization of soil survey efforts in similar areas.