Identification of Soil Erosion in Alpine Grasslands on High-Resolution Aerial Images: Switching from Object-based Image Analysis to Deep Learning?

Soil erosion in Alpine grassland areas is an ecological threat caused by the extreme topography, prevailing climate conditions and land-use practices but enhanced by climate change (e.g., heavy precipitation events, changing snow dynamics) in combination with changing land-use practices (e.g, more intensely used pastures). To increase our understanding of ongoing soil erosion processes in Alpine grasslands, there is a need to acquire detailed information on spatial extension and temporal trends.

In the past, we have successfully applied a semi-automatic method using an object-based image analysis (OBIA) framework with high-resolution aerial images (0.25-0.5m) and a digital terrain model (2m) to map erosion features in the Central Swiss Alps (Urseren Valley, Canton Uri, Switzerland). 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) (Zweifel et al. 2019). We now aim to apply a deep learning (DL) model with the purpose of fast and efficient spatial upscaling(e.g., alpine-wide analysis). While OBIA yields high quality results, there are multiple constraints, such as labor-intensive steps and the requirement of expert knowledge, which make the method unsuitable for larger scale applications. The results of OBIA are used as a training dataset for our DL model. The DL approach uses fully-convolutional networks with the U-Net architecture and is capable of rapid segmentation and classification to identify areas with reduced vegetation cover and bare soil sites.

Results for the Urseren Valley (Canton Uri, Switzerland) show an increase in total area affected by soil degradation of 156 ±18% during a 16-year observation period (2000-2016). A comparison of the two methods (OBIA and DL) shows that DL results for the Urseren Valley follow similar trends for the 16-year period and that the segmentations of eroded sites are in good agreement (IoU = 0.83). First transferability tests to other valleys not considered during training of the DL model are very promising, confirming that DL is a well-suited and efficient method for future projects to map and assess soil erosion processes in grassland areas at regional scales.

 

References

L. Zweifel, K. Meusburger, and C. Alewell. Spatio-temporal pattern of soil degradation in a Swiss Alpine grassland catchment. Remote Sensing of Environment, 235, 2019.