Mapping Swedish Soils with High-resolution DEM-derived Indices and Machine Learning

There is a soaring demand for up-to-date and spatially-explicit soil information to address various environmental challenges. One of the most basic pieces of information, essential for research and decision-making in multiple disciplines is soil classification. Conventional soil maps are often low in spatial resolution and lack the complexity to be practical for hands-on use. Digital Soil Mapping (DSM) has emerged as an efficient alternative for its reproducibility, updatablity, accuracy, and cost-effectiveness, as well as the ability to quantify uncertainties.

Despite DSM’s growing popularity and increasingly wider areas of application, soil information is still rare in forested areas and remote regions, and the integration with high-resolution data on a country scale remains limited. In Sweden, quaternary deposit maps created by the Geological Survey of Sweden (SGU) have been the main reference input for soil-related research and operation, though most parts of the country still warrant higher quality representation. This study utilizes machine learning to produce a high-resolution surficial deposits map with nationwide coverage, capable of supporting research and decision-making. More specifically, it: i) compares the performance of two tree-based ensemble machine learning models, Extreme Gradient Boosting and Random Forest, in predictive mapping of soils across the entire country of Sweden; ii) determines the best model for spatial prediction of soil classes and estimates the associated uncertainty of the inferred map; iii) discusses the advantages and limitations of this approach, and iv) outputs a map product of soil classes at 2-m resolution. Similar attempts around the globe have shown promising results, though at coarser resolutions and/or of smaller geographical extent. The main assumptions behind this study are: i) terrain indices derived from digital elevation model (DEM) are useful predictors of soil type, though different classification algorithms differ in their effectiveness; ii) machine learning can capture major soil classes that cover most of Sweden, but expert geological and pedological knowledge is required when identifying rare soil types.

To achieve this, approximately 850,000 labeled soil points extracted from the most accurate SGU maps will be combined with a stack of 12 LiDAR DEM-derived topographic and hydrological indices and 4 environmental datasets. Uncertainty estimates of the overall model and for each soil class will be presented. An independent dataset obtained from the Swedish National Forest Soil Inventory will be used to assess the accuracy of the machine learning model. The presentation will cover the method, data handling, and some promising preliminary results.