SoilGrids produces maps of soil properties for the entire globe at medium spatial resolution (250 m cell size) using state-of-the-art machine learning methods to generate the necessary models. It takes as inputs soil observations from about 240 000 locations worldwide and over 400 global environmental covariates describing vegetation, terrain morphology, climate, geology and hydrology. The aim of this work was the production of global maps of soil properties, with cross-validation, hyper-parameter selection and quantification of spatially explicit uncertainty, as implemented in the SoilGrids version 2.0 product incorporating state-of-the-art practices and adapting them for global digital soil mapping with legacy data.

The paper presents the evaluation of the global predictions produced for soil organic carbon content, total nitrogen, coarse fragments, pH (water), cation exchange capacity, bulk density and texture fractions at six standard depths (up to 200 cm). The quantitative evaluation showed metrics in line with previous global, continental and large-region studies. The qualitative evaluation showed that coarse-scale patterns are well reproduced. The spatial uncertainty at global scale highlighted the need for more soil observations, especially in high-latitude regions.

This study represents a considerable effort to provide a globally consistent product using the point dataset available to ISRIC, a large number of relevant covariates and some optimisation of a well-established machine learning method, within the limits of practical computation. Yet it is clear that this product has some limitations, which will be considered in further work.

First, there is an ever-expanding group of new covariates that can help explain and model the spatial variation of soil properties. Products derived from Earth observation are particularly relevant in this regard and have considerably improved over the last decade. For example, the European Space Agency Sentinel missions (both optical and radar) provide high-resolution data that have been shown to improve DSM model performances.

Second, a fundamental problem is a lack of well distributed point observations within the soil property geographic and features space. Additional soil data for so far under-represented regions, for example the northern boreal regions as being collated by the International Soil Carbon Network (Malhotra et al., 2019), will be sought for possible consideration in the WoSIS workflow that provides the point data underpinning the SoilGrids mapping effort. This effort would be aided by the provision by more data providers of at least a representative part of their point data to WoSIS, under suitable license. It is also important to consider the distribution of the observations in the covariate space to minimise the issues related to predictions into unknown regions of feature space (Meyer and Pebesma, 2020).

This study presents and discusses the production of global maps of soil properties as implemented in the SoilGrids 2.0 product, with cross-validation, hyper-parameter selection and quantification of uncertainty, using the best available (shared) soil profile data for the world. In particular, the study describes a robust and reproducible DSM workflow addressing the challenges of global data modelling:

non-homogeneous spatial distribution of input soil observations;

robust quantitative evaluation with a cross-validation procedure balancing accuracy and performances;

qualitative evaluation of the spatial patterns of the maps to include information about matching with well recognised pedo-landscape features;

quantification and mapping of the spatial uncertainty to provide users with a measure for and warning for users of the products.

As such, it describes a next step into global modelling and mapping of soil properties, explicitly highlighting the importance of quantitative and qualitative evaluation and uncertainty communication. The actual use of SoilGrids 2.0 in global and wide-area regional applications, where soil properties are important model inputs, will be the real test of its applicability and usefulness.