National-Scale Multi-Depth Digital Soil Mapping in Namibia Based on Legacy Data

Reliable national-scale, multi-depth soil property maps remain scarce across much of Africa, particularly in arid regions. Here, we present the first national-scale digital soil mapping (DSM) for Namibia: sixteen physical and chemical soil properties predicted on a 90 m prediction grid for three depth intervals (0–30, 30–60 and 60–100 cm). The framework draws on 4958 legacy profiles and augerings and 65 covariates, including locally produced datasets that capture soil–environment relationships not fully represented by global covariates. A semi-automated workflow combined Random Forest modelling in Google Earth Engine with R-based depth harmonisation, Boruta-feature selection and hyperparameter tuning. Performance and bootstrap variability were quantified over 20 bootstrap iterations per property–depth combination, with pedological evaluation and independent validation. pH was the most consistently predicted property, and base saturation, sand and silt captured useful broad spatial patterns, whereas bulk density, organic carbon, phosphorus, magnesium and clay were less predictable, and electrical conductivity and sodium showed low predictive skill. Performance generally declined with depth, and bootstrap variability was highest in sparsely sampled regions and deeper layers. Against independent samples from arable land, root mean square error was lower than SoilGrids in all 14 property–depth combinations evaluated and lower than iSDA in 8 of 12. Predicted patterns were consistent with known Namibian soil–landscape relationships and pedogenic processes, except for sodium. National-scale DSM is therefore feasible in data-sparse arid environments when harmonised legacy data are combined with global and locally relevant covariates, using simple but operationally robust methods. The products are intended for national and regional assessment, rather than site-specific decisions, and the modular workflow supports future updates and is potentially adaptable to other data-limited regions.