Mapping soil depth in southern pampas Argentina using ancillary data and statistical learning

Abstract

Restrictive layers such as hardpans limit the soil water and nutrients available for crops. In the southern Argentinean pampas, petrocalcic hardpans are found at variable depth within the field. Mapping the spatial distribution of depth to the petrocalcic hardpan is important for proper land evaluation, use, and management. Intensive grid sampling and spatial interpolation are labor-intensive and time-consuming mapping approaches for this soil property. In addition, the spatial distribution of this property is often difficult for interpolation techniques such as kriging. This study's objective was to evaluate the potential of soil electrical conductivity (ECa) and terrain attributes to map within-field spatial variation of soil depth using statistical learning techniques. Soil depth measurements up to 1-m depth were taken in eight fields at spatial sampling density ranging from 2 to 12 points ha -1 . Spatially dense data of elevation, terrain attributes, and ECa were migrated to soil depth sampling points and also spatially aggregated to include spatial information into the featured space. Then, random forest regression models were used to predict soil depth from colocated ECa and terrain data. Models were cross-validated using a k-fold approach using entire fields as folds. The overall model (using all data) resulted in out-of-the bag R 2 and RMSE of .66 and 22.8 cm respectively. Shallow (0-30 cm) and deep (0-90 cm) ECa values were the most important variables, accounting for variability at different ranges, but the importance varied between the soil types. These results suggested that field-scale ECa data and terrain attributes have potential to predict soil depth to hardpan. Further research is needed to improve the generalization of these models and improve the representation of spatial effects in order to implement site-specific management based on soil depth maps.