Digital soil mapping via machine learning of agronomic properties for the full soil profile at within-field resolution

Abstract

Fine-resolution maps of agronomic soil properties are essential for capturing within-field variability, supporting precision agriculture, improving understanding of soil–crop interactions, and providing reliable inputs for agroecosystem models. This study evaluated the use of digital soil mapping (DSM) with machine learning to predict 18 properties to a depth of 200 cm. Prediction performance peaked at shallow subsurface depths (15–30 cm), where the influence of dynamic anthropogenic disturbances diminished, and the relationship with processes captured by remote sensing remained strong. Total nitrogen, total organic carbon, and calcium showed the highest accuracy for surface depths (<30 cm) with model efficiency coefficient (MEC) of 0.68–0.79, while sand, clay, and K at mid-depths (30–60 cm) exhibited reasonable accuracy (MECs 0.42–0.5). About 17% of models performed worse than the observed mean baseline. Particle size fraction models showed reduced accuracy at the surface, likely due to episodic surficial processes like erosion. However, performance improved in mid-depths and decreased at greater depths due to lithologic discontinuities. While most models’ MEC declined with depth, root mean squared error remained low due to the homogeneity of parent material. This suggests low spatial accuracy may be acceptable if error across all locations is minimal, which is more important for applications that require minimized error propagation (e.g., crop modeling). Covariate importance analysis showed terrain variables remained predictive at greater depths, while surface imagery became less informative. Trend analysis by hillslope position demonstrated DSM's ability to capture site differences, such as the divergence of topographic patterns with different land management practices.