A First Approximation of the Global Soil Map for Europe.

The origin of the GlobalSoilMap.net (GSM) project dates back to 2006, its aim is mapping soils and their properties over the whole world with a high resolution using digital soil mapping techniques. The European Commission joined the GSM consortium through its Joint Research Centre Directorate, agreeing to deliver a GSM product by 2015.

In this presentation we will discuss the challenges and the methodologies adopted to create the first approximation of the GMS products for Europe. The product is still in its infancy and should be considered as a beta version of the final maps. Given its early stage, European countries are strongly encouraged in amending, integrating and validating this product with their own data. At the same time the development of this beta product was useful to tune methodologies and test different types of mapping procedures.

In this stage the GSM product was created using a bottom down approach, starting from a series of datasets covering the whole extent of the European Union. This approach facilitates the production stage as it minimizes the necessity of harmonizing data among countries. Moreover the mapping was based as much as possible on remotely sensed covariates reducing the use of legacy soil data to a minimum. This approach also reduces the necessity of harmonizing and avoids the issues connected with different mapping approaches in different countries.

In spite of this seemingly simplistic approach, the prediction performance of the models fitted (as estimated using cross-validation) is quite high for many of the mapped soil properties, reaching R² values of 0.7 for the estimation of topsoil pH. However, other soil properties were more difficult to predict as Soil Organic Carbon (R² of 0.39) and some properties had to be derived from pedotransfer functions (i.e. Bulk density).

Two main challenges will also be discussed. One is how to cross-predict soil properties between different databases on the basis of their overlapping information content and their proximity in environmental features space, to impute missing soil data (i.e. when a database lacks several chemical or physical properties or contains data about just the topsoil). The second is the estimation of uncertainty when Machine Learning (ML) techniques are applied, where the estimation of variance is not straightforward.