Predictive mapping of soil textural classes – Digital soil mapping case study across formerly glaciated terrains

In countries of northern Europe, formerly covered by continental glaciers, soil textural maps are a fundamental source of information for authorities, researchers and non-governmental organizations in the land use sector, environmental conservation and land use planning.  Yet, the small-scale mapping is still incomplete. Digital Soil Mapping (DSM) may provide ways to subjectively and efficiently generalize a low number of field observations into regional or country-wide soil textural maps by utilizing existing DEMs, remote sensing and airborne geophysical data.

We followed a commonly applied national soil textural classification, called RT (9 classes: 1 unsorted, 4 sorted, bedrock outcrops, stones and boulders, peat). Two mapping depths were selected: surface sediment (40‒90 cm) and base sediment (> 90 m). Three study areas (9 km to 18 km in width) covering wide geological variation were selected across the country based on the availability of Surficial deposits maps at 1:20/50 000 scale, and a 2-m-DEM and airborne geophysical datasets with 50‒75 m line spacing and four frequencies of electromagnetic data. The latter raster data were complemented by DEM derivatives, canopy height model (CHM), optical Sentinel-2 and ALOS PalSAR data resulting in a dataset of 253 explanatory features. Because the number field observations of soil texture were low for training and testing of machine learning models, the study areas were combined into one dataset (i.e. field data, surface sediment n=5133, base sediment n=4009). In addition, a complementary ‘pseudo’ reference dataset was extracted from the existing maps with GIS operations (i.e. map data, surface sediment n=9817, base sediment n=10120). We applied supervised classification with random forest (RF), and performed a priori feature selection with genetic algorithm and feature importance evaluation with permutation feature importance.

The overall classification accuracies for the surface sediment classification based on field data was 85.5%, and based on map data 78.1%. For the base sediment classification the respective overall accuracies were 74.6% and 75.6%. Class-specific accuracies were highest for the most common classes with abundant training data, while classes with fewer samples were poorly classified. For example, field data-based analysis revealed that bedrock outcrops and peat achieved the highest accuracies, while sandy till, fine sand, coarse silt, fine silt, and clay, represented by the lowest number of training data, consistently showed lower performance. The feature importance results indicate that DEM and its textural derivatives were the most significant features. However, optical and SAR satellite and radiometric airborne data would also be required for best separation of the soil textural classes if the classification was to be applied across wider areas.

Although these predictive mapping results indicate moderately successful surface sediment and base sediment textural classification, the small size of the study areas and consequent highly unbalanced training sets, poses a limitation for generalization of this study. However, it indicates a potential for successful surface sediment and base sediment textural classification with DSM if applied at regional scale or country-wide. Future studies should explore the applicability of supervised deep learning algorithms to avoid calculating a high number of textural derivatives of the DEM.