Using Machine Learning Methods to Predict Water Erosion Patterns in Northern Germany

Recent studies show that machine learning methods (ML) have a high potential to model various soil erosion processes. The main focus of this research is on qualitative modelling (risk classes, binary occurrence) on smaller scales e.g., probability maps for gully erosion. In this study, we used four machine learning methods to reproduce spatial explicit information on soil loss rates in a 5 m resolution obtained in the Lower Saxonian (Germany) soil erosion monitoring program in the years 2000 - 2020. The monitoring data includes information on soil loss by water erosion (sheet and rill erosion) on 465 ha of cropland in northern Germany derived by continuous erosion feature mapping after erosive rainfall events (see Steinhoff-Knopp & Burkhard (2018) for methods and results). We applied the ML methods Random Forest (RF), a Single-Layer Neural Network (SLNN), a Deep Neural Network with multiple hidden layers (DNN) and a Convolutional Neural Network (CNN) to reproduce the mapped soil loss rates. Prediction variables included are up to 19 soil, land use, rainfall and DEM-derived topographic parameters. All ML methods were able to reproduce the soil erosion patterns. The comparison between the different models shows that the CNN model outperforms all other tested models in nearly all metrics. Its RMSE of 1.05 and MAE of 0.41 are significantly lower than those of the RF (RMSE: 1.31, MAE: 0.58) and SLNN (RMSE: 1.48, MAE: 0.63). Only the DNN performs similarly, with a slightly higher RMSE of 1.1 and MAE of 0.58. However, the classification performance of the RF, DNN, and CNN models is comparable, with F1 scores ranging from 0.68 to 0.70 and AUC values between 0.92 and 0.94. Additionally, the permutation importance was calculated to assess the influence of the predictor variables. In all four models, the variable with the highest importance is the DEM. Its importance ranges from 15% to 18.3%, depending on the model. All models also strongly rely on USLE C and R factors. Our findings emphasize the high potential of ML-driven erosion predictions and will be rolled out to predict soil erosion rates on cropland in northern Germany.

Steinhoff-Knopp, B. and Burkhard, B.: Soil erosion by water in Northern Germany: long-term monitoring results from Lower Saxony, CATENA, 165, 299–309, https://doi.org/10.1016/j.catena.2018.02.017, 2018.