Optimizing Soil Aggregate Stability Predictions with Machine Learning: A Comparative Analysis of Input Variables and spatial transferability

Land degradation have become critical environmental issues, leading to reduced soil and water quality and reduced yields in arable lands. Soil aggregate stability (AS) refers to the ability of soil aggregates to resist disintegration or breakdown and thus presents a measure to assess the soils susceptibility to applied forces. Despite its significance, for example in soil erosion processes, spatial data on AS is scarce and only few studies on the spatial behaviour of AS exist, frequently due to the high monetary, work and time expense needed to gain data. 

Machine learning approaches are increasingly used due to their high accuracy when incorporating various co-variables and are already achieving promising results in the field of AS. However, it is often unclear how well these models perform in different environmental conditions and landscapes, particularly outside their training sites.

The aim of this study is to identify and compare the best-performing variables for the soils in two study sites in northern Germany with different environmental conditions and to evaluate if and how well a model, trained on one study site would perform on another. To accomplish this, a total of 500 topsoil samples (250 each) were collected from the two study sites. They were analysed for soil properties, including AS, soil texture, SOC, pH, electrical conductivity. Additionally, a range of topographic indices and additional data (e.g. Crop, geology)  were analysed.

The preliminary results show, that SOC, topographic wetness index (TWI), slope and the fine sand fraction were deemed the best performing variables in the random forest model in study site A. The model achieved an r2 of 0.575 and RMSE 7.992. Analysis of the terrain indices in study site B show channel network base level, aspect and analytical hillshading as the best performing terrain indices. A performance gap of the model would indicate limited transferability, as the model may have overfitted to site A's specific landscape and conditions.  Additional models will be tested to determine which ones transfer more effectively between different sites.