Abstract:
Soil erosion by water remains one of the most pressing environmental challenges for sustainable land management, agricultural productivity, and water quality protection. In Central Europe, and particularly in Germany, the problem is becoming more urgent as climate change intensifies rainfall extremes, agricultural landscapes become more homogenized, and traditional empirical standards increasingly fail to represent contemporary erosive conditions. Accurate assessment of rainfall erosivity, reliable simulation of sediment transport, and rapid identification of erosion- and flood-affected areas are therefore essential for evidence-based soil conservation planning. This dissertation addresses these challenges by integrating process-based modeling, deep learning, and remote sensing into a unified framework for erosion assessment and environmental monitoring in Germany.
The first part of the dissertation focuses on rainfall erosivity, the climatic driver of water erosion. Conventional approaches in Germany, in the absence of high-resolution precipitation data, are based on empirical relationships derived from historical precipitation conditions and do not adequately capture recent shifts in rainfall intensity. To address this limitation, a deep learning-based framework was developed to reconstruct historical rainfall erosivity and project future changes across Germany at high spatial resolution. By combining RADKLIM radar precipitation data with a Hypernetwork-Enhanced CNN model for historical reconstruction and a ConvLSTM encoder-decoder for near-term projection, the study produced a continuous century-scale erosivity dataset from 1931 to 2043. The results revealed a clear long-term increase in erosive forcing. This work demonstrates that data-driven models can serve as effective alternatives for estimating and forecasting rainfall erosivity in the absence of high temporal resolution data.
The second part of the dissertation evaluates the performance of the spatially distributed erosion and sediment transport model WaTEM/SEDEM under soil conservation conditions. Six highly instrumented micro-scale watersheds in Southern Germany, monitored over eight years, were used to test the model’s ability to reproduce low sediment yields in landscapes characterized by conservation-oriented management and structural sediment control measures. A Generalized Likelihood Uncertainty Estimation framework (GLUE) was applied to assess parameter uncertainty and model plausibility across different spatiotemporal scales. The results show that WaTEM/SEDEM can reproduce the approximate magnitude of long-term sediment yields, but its performance is limited at annual and micro-watershed scales, especially where sediment fluxes are very low and landscape structures strongly influence connectivity. These findings highlight the importance of scale-aware model evaluation and demonstrate that model suitability depends not only on statistical fit, but also on the intended management application.
The third part of the dissertation addresses the challenge of automated environmental monitoring through deep learning-based image segmentation. Two related studies were conducted: the first developed Erosion-SAM, a fine-tuned adaptation of the Segment Anything Model for delineating water erosion and deposition features in high-resolution aerial imagery; the second compared fine-tuned SAM and U-Net architectures with ResNet backbones for rapid flood mapping from UAV and helicopter imagery. Both studies show that transfer learning and foundation models can substantially reduce the need for large labeled datasets while maintaining strong segmentation performance. For erosion mapping, the adapted SAM successfully identified visually heterogeneous erosion structures across different land cover types. For flood mapping, both SAM-based and U-Net-based models achieved high accuracy, with strong practical potential for rapid post-event assessment and disaster response. Together, these studies demonstrate the value of modern computer vision methods for generating spatially explicit environmental observations that can support both scientific analysis and operational decision-making.
Taken together, the dissertation presents an integrated approach that links climatic forcing, process-based erosion modeling, and automated image-based monitoring. The overarching contribution lies in showing how deep learning can be used not as a replacement for environmental theory, but as a powerful complement to classical soil erosion science. The results provide improved tools for rainfall erosivity estimation, a critical evaluation of model performance under conservation conditions, and new methods for the rapid mapping of erosion and flood impacts. In doing so, the dissertation contributes to a more adaptive, data-rich, and spatially explicit framework for soil conservation planning, climate change impact assessment, and natural hazard management.