Digital Soil Mapping — Predictive Soil-Landscape Modelling
Digital Soil Mapping · Also known as: DSM, predictive soil mapping, quantitative soil-landscape modelling, geostatistical soil mapping
Digital Soil Mapping (DSM) is a quantitative, data-driven pipeline that predicts the spatial distribution of soil properties and classes across a landscape by statistically linking field observations to environmental covariates — terrain attributes, remote sensing imagery, climate surfaces, and geology layers. The approach replaces or augments traditional expert-drawn soil surveys with reproducible, spatially explicit models, and is applied in agronomy, land management, food security, and environmental assessment.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
DSM is appropriate when spatially continuous estimates of soil properties or soil classes are needed across an area too large or heterogeneous to map exhaustively by conventional profile-and-transect survey. It suits precision agriculture (variable-rate input management), national and regional soil inventories, carbon accounting, land degradation assessment, and hydrological modelling. The method requires field soil observations linked to geographic coordinates and at least one layer of environmental covariates. It is not appropriate when only a handful of point observations are available without spatial covariates, or when the mapping target is a highly localised anthropogenic soil feature (e.g., contaminated hotspot) with no environmental signal in the available covariates.
Strengths & limitations
- Produces spatially continuous, quantitative soil property maps with explicit uncertainty estimates — a major advance over polygon-based conventional maps.
- Reproducible and updatable: the entire workflow from covariates to output map can be scripted and rerun as new observations or better covariates become available.
- Scales from field to national level using the same algorithmic framework, with covariate resolution adapted to the mapping scale.
- Integrates diverse data sources — field surveys, legacy profiles, remote sensing, terrain analysis — within a single modelling framework.
- Machine-learning models such as random forest handle nonlinear soil-landscape relationships and high-dimensional covariate sets without requiring explicit parametric assumptions.
- Prediction accuracy is constrained by the quality and spatial coverage of the training sample; poorly designed or sparse sampling leads to biased maps regardless of model sophistication.
- Environmental covariates must correlate with the target soil property; if the dominant soil-forming factor is not captured in available layers (e.g., micro-scale management history), model explanatory power will be low.
- Spatial autocorrelation in training data inflates cross-validation accuracy when random rather than spatial splits are used, leading to overoptimistic performance reports.
- High-resolution DSM is computationally demanding; processing continental or global covariate stacks at fine resolution requires substantial computing infrastructure.
- Uncertainty maps are only as reliable as the model and validation sample; they do not capture model structural uncertainty or covariate measurement error.
Frequently asked
How many soil samples do I need for DSM?
There is no universal minimum, but rules of thumb in the literature suggest at least 50–100 observations for simple landscapes and single properties, with several hundred preferred for complex terrain or multiple targets. Conditioned Latin hypercube sampling (cLHS) is widely used to select a compact sample that spans the multivariate covariate space efficiently. Sample size requirements also depend on model complexity: random forests tolerate smaller samples better than deep learning approaches.
Which model algorithm works best for DSM?
No single algorithm dominates across all soils, landscapes, and target properties. Random forest and gradient boosting (XGBoost, LightGBM) are currently the most widely used because they handle nonlinearity and high-dimensional covariates without distributional assumptions and provide variable importance measures. Regression kriging remains competitive when spatial autocorrelation in the residuals is strong. Comparing two or three algorithms with spatial cross-validation and selecting based on RMSE or Lin's CCC on the validation set is good practice.
What is SCORPAN?
SCORPAN is the mnemonic for the seven soil-forming factors used as the conceptual basis for covariate selection in DSM: Soil (existing soil data), Climate, Organisms (vegetation, land use), Relief (terrain), Parent material, Age (time), and spatial position (N). It extends Jenny's classical CLORPT model to include available soil information and geographic coordinates as predictors, framing DSM as a spatial interpolation problem anchored in pedological theory.
Can DSM replace traditional soil survey?
DSM complements rather than replaces expert field knowledge. Conventional soil survey provides the observations and pedological interpretation that anchor DSM models; without field data there is nothing to train the model on. DSM adds spatial continuity, reproducibility, and quantitative uncertainty that polygon maps lack. The two approaches are increasingly combined: expert surveyors guide sampling design and interpret outputs, while DSM handles spatial prediction and map updating.
Sources
- McBratney, A. B., Mendonca Santos, M. L., & Minasny, B. (2003). On digital soil mapping. Geoderma, 117(1–2), 3–52. DOI: 10.1016/S0016-7061(03)00223-4 ↗
- Minasny, B., & McBratney, A. B. (2016). Digital soil mapping: A brief history and some lessons. Geoderma, 264, 301–311. DOI: 10.1016/j.geoderma.2015.07.017 ↗
How to cite this page
ScholarGate. (2026, June 3). Digital Soil Mapping. ScholarGate. https://scholargate.app/en/agronomy/digital-soil-mapping
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Random ForestMachine learning↔ compare