Mapping soil structure across Switzerland using machine learning

Soil structure class maps were developed for five standard depth intervals (0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm, and 60–100 cm) using a Random Forest machine learning approach at 30 m spatial resolution. The modelling dataset was derived from the NABODAT soil database version 6 (Service Centre NABODAT, 2022) and consisted of approximately 13,000 soil observations distributed across Switzerland. Observations were primarily concentrated in the Central Plateau and Jura regions, while comparatively few samples were available from the Alpine region. As a result, prediction uncertainties are expected to be higher in under-sampled areas, particularly in the Alps. The objective of this work was to provide a first national-scale assessment of soil structure class distribution. The NABODAT database contains multiple soil structure classes, including granular, subangular blocky, angular blocky, massive, prismatic, and single grain. Soil properties including clay content (%), sand content (%), and soil organic carbon (SOC, %) were extracted for each soil profile. To harmonize soil properties to standard depth intervals, equal-area spline functions implemented in the R package ithir (function ea_spline) were applied to clay, sand, and SOC observations. This procedure generated continuous estimates for each target depth interval. Soil structure classes were assigned to standard depth intervals using a dominant overlap approach. For each profile, the overlap between observed soil horizons and the target depth interval was calculated. The soil structure class with the highest cumulative overlap within each interval was assigned as the representative class. For example, if a profile contained a granular horizon from 0–2 cm and a subangular blocky horizon from 2–6 cm, then the 0–5 cm interval would include 2 cm of granular soil and 3 cm of subangular blocky soil, resulting in classification as subangular blocky. In rare cases where two classes had equal overlap, the class occurring closer to the soil surface was selected. After depth harmonization, environmental covariates were extracted for each observation location. These included climatic, topographic, terrain, spectral, vegetation, and soil-related variables. Initially, approximately 77 covariates were considered. Random Forest models were fitted using the ranger package in R, and model performance was evaluated using 5-fold cross-validation. Permutation-based variable importance was computed for each fold, and the 20 most important covariates were selected for final model development. Final Random Forest models were trained separately for each depth interval. The models were then applied to national-scale environmental covariate layers to generate spatial predictions of soil structure classes across Switzerland. Class probabilities were derived from the Random Forest outputs. Prediction uncertainty was calculated as: Uncertainty = 1 − max(Pi) where max(Pi) represents the highest predicted class probability among all soil structure classes at a given location. Lower values indicate higher prediction confidence, while higher values indicate greater uncertainty. The covariates used in this study are described in Gupta et al. (2026). Model performance for each depth interval is summarised below: Accuracy Kappa 0-5 0.73 0.42 5-15 0.71 0.4 15-30 0.6 0.41 30-60 0.75 0.31 60-100 0.73 0.44 Model performance varied across depth intervals, with overall accuracies ranging from 0.60 to 0.75. The relatively lower Kappa values (0.31–0.44) suggest moderate agreement, likely influenced by class imbalance in the soil structure dataset. The model showed reduced predictive performance for minority soil structure classes, which is consistent with the observed class imbalance in the dataset and is reflected in the moderate Kappa values. Class Encoding For the 0–5 cm, 5–15 cm, and 15–30 cm depth intervals: 0 = No Data (NA); 1 = Angular blocky; 2 = Cloddy; 3 = Granular; 4 = Massive; 5 = Platy; 6 = Prismatic; 7 = Single grain; 8 = Subangular blocky For the 30–60 cm and 60–100 cm depth intervals: 0 = No Data (NA); 1 = Angular blocky; 2 = Granular; 3 = Massive; 4 = Platy; 5 = Prismatic; 6 = Single grain; 7 = Subangular blocky Important Note The cloddy and platy structure classes were represented by relatively few observations in the training dataset. Consequently, model performance for these classes was limited, and predictions for these classes should be interpreted with caution. Data and code availability Spatial prediction maps, class probability layers, and uncertainty maps for all depth intervals are provided in this Zenodo repository. The code used to develop these maps is available at: https://github.com/ETHZ-repositories/Soil_structure_maps Other related soil property maps for Switzerland, including soil organic carbon (SOC), clay, sand, nitrogen, and phosphorus, are available at: https://doi.org/10.5281/zenodo.7821649 Bulk density maps are available at: https://doi.org/10.5281/zenodo.15274427 Coarse fragment maps are available at: https://doi.org/10.5281/zenodo.17453998 Reference: Gupta, S., Scheper, S., & Alewell, C. (2026). Mapping Swiss soil bulk density at 30 m Resolution: Insights from Machine Learning, environmental Covariates, and national data. International Journal of Applied Earth Observation and Geoinformation, 146, 105112. Service Centre NABODAT., 2022. Swiss Soil Dataset – Documentation Version 6 (April 2022), https://ccsols.ch/wp-content/uploads/2023/11/DOCUMENTATION_V6_2022_EN.pdf.