Long-Term Trends in Rainfall Erosivity Derived from a Deep Learning Reconstruction Framework

Quantifying long-term soil erosion dynamics, for example, to better assess the effects of climate change on soil erosion, requires temporally continuous records of rainfall erosivity, a key input for explicit soil loss modeling. However, high-temporal-resolution precipitation data are rarely available historically at large scales. We present a data-driven framework to overcome this constraint and to reconstruct annual gridded rainfall erosivity (R-factor) for Germany from 1930 to 2000. First, high-temporal-resolution gridded precipitation from the RADOLAN product (2001–2025) was used to compute spatial maps of rainfall erosivity for the modern period based on standard intensity-based erosivity formulations. These modern R-factor maps served as target fields to train a convolutional neural network (CNN) that learns the relationship between erosivity and commonly available predictors, including monthly and annual precipitation statistics, temperature indices, and large-scale climatic indicators. The CNN model was trained and evaluated using spatial–temporal cross-validation and independent station holdouts to quantify predictive skill and uncertainty. After validation, the model was applied retrospectively to coarse-resolution historical climate records to generate annual 1-km gridded R-factor estimates for 1930–2000. The reconstructed time series reveal spatially coherent patterns and multi-decadal trends in rainfall erosivity that are not captured by coarse aggregated proxies, and they provide a physically informed dataset for retrospective soil erosion modeling and climate-impact assessments.