Using machine learning to model Soil Aggregate Stability as an indicator for soil erosion susceptibility at the catchment scale

Soil aggregate stability (AS) is a key component for numerous soil processes and a significant factor in soil erosion. Despite its significance, the research on AS has been comparatively limited, possibly due to the high monetary, work and time expense needed to gain data. Furthermore, the spatial distribution of aggregate stability is influenced by various topographic and physical factors, such as surface curvature and soil characteristics. It is also affected by non-numerical conditions, for instance land use and crop type. This combination of quantitative and qualitative variables highlights the complexity of AS modeling. Therefore, it is even more important to gain further insight on the spatial distribution and prediction techniques suitable for AS.

The aim of the ESTABLE project is to model the spatial variability and distribution of AS and analyze its relationship to soil erosion processes at the catchment scale. To accomplish this, a total of 500 topsoil samples were collected from the two study sites in Northern Germany (Lamspringe and Ascheberg). All soil samples were analyzed for aggregate stability, soil texture, organic carbon content, pH, and electric conductivity.

To represent the complexity of the relationship of factors influencing AS, various machine learning models, including Boosted Tree and Random Forest, are tested to implement categorical data in addition to the wide range of numerical input variables. These models were evaluated based on their performance, parameterization, and interpretability in comparison to traditional interpolation techniques like multiple linear regression and regression kriging. It has become evident that most machine learning techniques are more effective at capturing the intricate interactions that influence aggregate stability.

The best performing model is then used to verify, that low aggregate stability areas are also prone to erosion. The use of UAVs and field mapping enable a detailed and accurate assessment of the spatial distribution of soil erosion. This model could also serve as a valuable tool for other sites and subsequent studies.