Comment on essd-2021-437

Abstract

Texture is a fundamental soil property for multiple applications in environmental and earth sciences. Knowing its spatial distribution allows for a better understanding of the response of soil conditions to changes in the environment, such as land use. This paper describes the technical development of Colombia's first texture maps, obtained via spatial ensemble of national and global digital soil mapping products. This work compiles a new database with 4203 soil profiles, which were harmonized at five standard depths (5, 15, 30, 60, and 100 cm) and standardized with additive log-ratio (ALR) transformation. A stack with 83 covariates was developed, including both quantitative and qualitative covariates, and harmonized at 1 square kilometer of spatial resolution. The top explanatory covariates were selected for each transformation in all standard depths through a recursive feature elimination. Ensemble Machine Learning (EML) algorithms (MACHISPLIN and landmap) were trained to predict the distribution of soil particle fractions (SPF) (clay, sand, and silt), and a comparison with SoilGrids (SG) products was performed. Finally, a spatial ensemble function was created, which identified the fewest prediction errors between EML and SG, then selected the better of these algorithms for each pixel and standard depth. The results of EML algorithms show that the accuracy of MACHISPLIN and landmap were very similar in the SPF at each standard depth, and both were more accurate than SG. The amount of variance explained (AVE) was between 0.12 and 0.35 for EML, and -0.17 and -0.01 for SG; the concordance correlation coefficient (CCC) was between 0.32 and 0.54 for EML, and 0.04 and 0.16 for SG. The best EML performance was found for the two superficial layers (5 and 15 cm). The accuracy of the spatial ensemble was higher compared to the other algorithms at all standard depths, but the largest improvement was found at the first layer, where AVE values increased between 0.04 and 0.13, and CCC values between 0.04 and 0.10. EML predictions were frequently selected for each PSF and depth in the total area; however, SG predictions were better when increasing soil depth in some specific regions such as Orinoquía and Amazonía. The final error distribution in the study area showed that sand fraction presented higher absolute error values than clay and silt fractions, specifically in eastern Colombia. The spatial distribution of soil texture in Colombia is a potential tool to provide information for water related applications, ecosystem services and agricultural and crop modeling. However, some aspects must be attended in future efforts to accurately map soil texture; for example, the treatment of abrupt changes in the texture between depths, unbalanced data, and compositional data consistency in spatial ensemble products. Our results and the compiled database (Varón-Ramírez and Araujo-Carrillo, 2021; Varón-Ramírez et al., 2021) provide new insights to solve some of the aforementioned issues.