Pedodiversity‐Based Covariates for Digital Soil Mapping Framework

ABSTRACT The integration of soil diversity as a covariate in digital soil mapping framework remains largely unexploited. However, we believe that it may carry important information that can be used to create more accurate digital maps of soil properties. Therefore, novel covariates are proposed based on the traditional concept of pedodiversity, which considers the number and abundance of individual soil classes, but exploited using the moving window (MW) method (an algorithm that proceeds incrementally and calculates statistics from individuals falling within a defined window). By moving a circular MW of a defined radius cell by cell across a rasterized (with a resolution of 50 m) Soil map of the Czech Republic (at a scale of 1:250,000), a diversity status value (Shannon's entropy) could be assigned to each raster cell. Furthermore, by varying the radius (4, 7, 10 and 15.2 km) of the MW, four different pedodiversity spatial patterns (PSP) were obtained. Subsequently, the PSPs were used together with other common covariates, such as morphometric terrain parameters (MTP) and climate, as covariates in the prediction of soil organic carbon (SOC) content (%) (depth 0–30 cm) for the forested land of the Czech Republic. To test the effect of the new covariates, several random forest models were calibrated (and validated) either using only the MTP and climate covariates or adding one of the PSPs (calculated with different MW radii). Furthermore, the difference between using the Soil map of the Czech Republic as a categorical covariate compared to PSP as numerical (continuous) one was assessed. As a result, the use of PSP led to a significant improvement in prediction accuracy, especially for PSP with the largest MW radius (15.2 km). The improvement was seen from 1.74 to 1.71 and from 0.46 to 0.49 in terms of RMSE and RSQ, respectively. PSP was shown to play an important role in improving model accuracy, as it (with a MW radius of 15.2 km) appeared among the top five covariates in terms of variable importance (generated with recursive feature elimination). Moreover, PSP as a numerical (continuous) covariate turns out to be more important than the polygon map of soil classes as a categorical one (the latter brought only a non‐significant improvement in prediction accuracy). The benefit of using PSP as a covariate was attributed to the existing relationship between soil classes and SOC content, while there is also a relationship between these soil units and the degree of pedodiversity.