Identifying Soil Erosion Processes in the Alps using Machine Learning Techniques

Soils are an important part of our ecosystem and a significant resource, which ought to be protected. A major threat to soil is the degradation caused by soil erosion. Alpine regions are affected by soil erosion mainly due to the prevailing climate conditions (triggered by wind, water/snow) and the steep terrain (gravitational processes). Soil erosion is further increased due to climate change related factors (e.g., more extreme precipitation events) as well as changing land-use practices (e.g., more intensely used/abandoned pastures). Monitoring of soil erosion processes is therefore crucial for gaining a holistic understanding of spatial and temporal developments. However, it is difficult to observe the extent of ongoing soil erosion in Alpine terrain due to the inaccessibility of regions, steep topography, and the vastness of affected regions. To successfully overcome many of these restrictions, we use remote sensing approaches with the use of machine learning techniques.
Our aim was to identify different soil erosion processes and to map the location and extent of these soil erosion sites. For this purpose, multiple mapping methods are explored. The results produced with the monitoring tools are further analysed to understand the spatial distribution and temporal developments of soil erosion sites and to identify potential causal factors. Degraded sites are classified according to the major erosion process (shallow landslides; sites with reduced vegetation cover affected by sheet erosion) or triggering factors (trampling by livestock; management effects).
In a first step, we developed a semi-automatic workflow based on Object-based Image Analysis (OBIA), with which soil erosion features are mapped on the basis of high-resolution aerial images (0.25-0.5 m, RGB spectral bands) and a digital terrain model (2 m). This approach is used to map erosion sites and assign corresponding classes located within the Urseren Valley (Central Swiss Alps, Canton Uri) at five different time steps during a 16-year study period (aerial images of 2000, 2004, 2010, 2013 and 2016). The area affected by soil erosion increased by a total of 156% ± 18% during these 16 years. OBIA yields high quality results but the workflow presents multiple constraints, such as labour and time intensive steps or a lack of transferability to other regions, which make the method unsuitable for larger scale applications.
We therefore applied a deep learning approach, which can be used in a faster and more efficient manner. This approach uses fully-convolutional neural networks with the U-Net architecture and is capable of rapid segmentation and classification of aerial images in order to identify soil erosion sites. The mapping results of the OBIA study are used as training data (9 km2 training area) for this U-Net mapping tool. We compare the results of the U-Net to those of OBIA for a held-out test region in the Urseren Valley (17 km2 testing area) and found that the U-Net performs on par with OBIA in terms of segmentation of erosion sites as well as the identified temporal trends (16-year period). Both the spatial (within the Urseren Valley) and temporal (data from new year not seen during training) transferability were tested. Due to method-specific differences, the U-Net achieves a F1-score of 78% when compared to OBIA results. However, visual assessments indicate that the U-Net is slightly more accurate than OBIA (U-Net maps additional affected sites).
With the U-Net tool, grassland areas can be mapped efficiently. Ten study sites across Switzerland (8 located in the Alps, 1 in the foothill region of the Alps, 1 in the Jura mountains) were selected to map the location and extent of shallow landslides, to identify potential causal factors and to better understand regional differences. Using a logistic regression with a Group Lasso variable selectionmethod we identify important variables from a set of explanatory variables consisting of traditional susceptibility modelling factors as well as climate-related factors representing local and cross-regional conditions. Due to different conditions of the study sites, different important explanatory variables are identified (regression model accuracies between 70.2 and 79.8%). However, slope and aspect are amongst the most consistent. A model evaluating all sites simultaneously (regression model accuracy 72.3%) confirms these findings and further selects topographic roughness as an additional important variable. Sites with better regression model performance are located in the Alps and tend to have an east-west orientation of the valley axis, possibly capturing processes related to slope exposition (e.g., snow gliding on South facing slopes).
As snow gliding is thought to be an important trigger for soil erosion during winter months, we calculate a spatial snow glide model (SSGM) for Switzerland based on a 30-year winter precipitation average. Modelled snow glide distances are compared to shallow landslide densities of the ten mapped study sites. Correlations were found between higher snow glide distances and higher shallow landslide densities (and vice versa). For more detailed evaluations on the connection between snow gliding and soil erosion, additional high resolution data sets (winter precipitation) and higher temporal resolutions of mapped erosion sites are needed, as snow gliding is generally event based (i.e., during strong ‘snow glide winters’).
During this thesis, a monitoring tool was developed, which allows for mapping and subsequent analysis of soil erosion sites located on grasslands with unprecedented efficiency. This will allow for future large scale applications such as nation-wide or even Alpine-wide studies. The produced data sets may accompany other studies on soil erosion (e.g., erosion modelling) and may serve as the basis for land-use mitigation strategies.