Towards the Development of Bias-Corrected Rainfall Erosivity Time Series for Europe

Matthews, F., Medved, A., Borrelli, P., Liakos, L., Verstraeten, G., Panagos, P., Bezak, N., 2025. Towards the Development of Bias-Corrected Rainfall Erosivity Time Series for Europe. Journal of Hydrology, 651, 132460. https://doi.org/10.1016/j.jhydrol.2024.132460

Rainfall erosivity maps for (near) real-time soil erosion predictions require the integration of (combinations of) reanalysis products and satellite retrievals of rainfall while overcoming potential biases related to their simplified spatial and temporal scales. Across Europe, we evaluate:

1. The European Meteorological Observations (EMO) dataset to simulate the localized characteristics of rainfall erosivity at the event scale (EI30).
2. Different implementations of quantile delta mapping (QDM) bias correction to improve prediction skill.

Between 1990 and 2014, evaluations were made at several spatial (location-specific, climatic zone, and pan-European) and temporal (event, annual, and long-term annual average) scales.

The uncorrected EMO predictions demonstrated:

• A slight overprediction of the number of EI30 events.
• A reduced coefficient of variation in EI30 (CV_EMO = 1.57; CV_REDES = 2.50).
• A relatively low location-specific predictive skill (R² = 0.22; n = 139,306), with larger discrepancies in Southern Europe.

Following QDM bias correction, the EI30 predictions better represented the large-sample variability of EI30 across climate regions and improved the monthly correspondence. At specific locations, station-wise bias correction was the only implementation to improve predictions at the:

• Event scale (R² = 0.24; n = 139,306),
• Annual scale (R² = 0.51; n = 14,248),
• Mean annual scale (R² = 0.76; n = 1,142).

While bias correction can improve rainfall erosivity predictions, applications should consider the potential for substantial error propagation into subsequent analyses, regional disparities in performance, and the possibility that methods improving large-sample statistical correspondence may simultaneously degrade location-specific time series predictions.