Spatially explicit prediction of soil aggregate stability based on machine learning approaches to identify soil quality degradation caused by soil erosion – ESTABLE –

该项目旨在对表土团聚体稳定性进行空间差异性预测。作为未覆盖或仅临时覆盖的土壤表面侵蚀敏感性的指标，团聚体稳定性的分布模式提供了与常规土壤侵蚀相关的土壤结构退化的指示。对于团聚体稳定性的空间表示，该项目开发了基于机器学习方法的区划模型，该模型不仅考虑了地貌面积变量和对团聚体稳定性的发展起决定性作用的地貌面积变量和泥沙运移，而且还考虑了不同类型的土壤耕作。拟开发的区划模型基于在数字土壤制图中广泛使用和检验的多元线性回归、回归克里格法、决策树(如回归树、增强回归树、增强决策树)和随机森林方法等学习方法。为了检验团聚体稳定性作为与侵蚀相关的土壤结构变化的指标的适宜性，并在定量可理解的量的基础上对其进行统计确保，区域化的团聚体稳定性与用侵蚀模型E3D(E3D)模拟30年的土壤侵蚀(净侵蚀量和侵蚀率)有关。空间上不同的净侵蚀量计算是基于该时间段内发生的所有侵蚀性降水事件。团聚体稳定性的空间格局与模拟的长期土壤侵蚀之间的关系最终用于建立一个回归模型，用于评价长期土壤侵蚀和侵蚀强度对团聚体稳定性空间分布格局的影响，以及土壤侵蚀敏感性的空间分异指示。

The project aims at the spatially differentiated prediction of the aggregate stability of the topsoil. As an indicator for the erosion susceptibility of uncovered or only temporarily covered soil surfaces, the distribution patterns of aggregate stability provide an indication of the degradation of soil structure associated with regular soil erosion. For the spatial representation of aggregate stability, the project develops regionalization models based on machine learning methods, which take into account not only the geomorphological area variables responsible for surface water and sediment transport and the physical and chemical soil properties which are decisive for the development of aggregate stability, but also different types of soil cultivation. The regionalization models that are to be developed are based on the learning methods such as multiple linear regression, regression kriging, decision tree (e.g. regression tree, boosted regression tree, boosted decision tree) and random forest methods that are widely used and tested in digital soil mapping. In order to test the suitability of the aggregate stability as an indicator for erosion-related soil structure changes and to statistically secure it on the basis of quantitatively comprehensible quantities, the regionalized aggregate stability is to be related to the soil erosion (net erosion quantities and rates) modelled for a period of 30 years with the erosion model EROSION-3D (E3D). The spatially different calculation of net erosion is based on all erosive precipitation events which occurred within the time period. The correlations between the spatial pattern of aggregate stability and the modelled long-term soil erosion finally serve to derive a regression model for the evaluation of the effects of long-term soil erosion and erosion intensities on the spatial distribution pattern of aggregate stability and for the spatially differentiated indication of soil erosion susceptibility.