Two domains vs single-domain ERA5 dynamical downscaling with WRF over Euro-Cordex improves precipitation hindcast: A 6-year case study over mainland France for agricultural studies

Increases in surface temperatures and changes in precipitation patterns due to climate change affect crop yields and require adaptation of agricultural systems. High-resolution climate data, especially precipitation, are critical for impact modelling in agriculture and difficult to obtain from general circulation models. Dynamical downscaling with regional climate models (RCM), such as the Weather and Research Forecasting (ARW/WRF) model, is widely used to generate such data. Despite their improvements, RCM rainfall simulations still contain biases that make it difficult, if not impossible, to use them directly in impact models. To address this, bias correction methods have been proposed to improve the performance of rainfall simulations, but they introduce additional sources of uncertainty (e.g. changes in the state of the climate regime) and remain controversial. These persistent problems in the RCM outputs are due to inherited biases in the forcing data, the limitations of the physical schemes and the downscaling protocol itself. The resolution and reliability of the ERA5 reanalyses lead us to compare one- and two-domain downscaling protocols to reproduce the local climate regime and variability over the main French agricultural production basins.

Both protocols share the Euro-Cordex geographical area as their first domain, while the second protocol adds another domain around France. The target grid cell resolution is 8 km. ERA5 reanalyses data forced the WRF parent domain every six hours along (1) the 1979-1985 period, (2) the yield-damaging summer drought of 2003, and (3) the low rainfall spring of 2011 for five agroclimatic zones in mainland France. Spectral nudging is applied only to the first domain, and subgrid-scale cloud-radiation interactions are activated. The study focuses on five agriculturally relevant variables: maximum and minimum temperatures (Tmax and Tmin), potential evapotranspiration (PET), and the annual amount and cycle of precipitation. These variables are critical for crop growth stages, irrigation management, and yield prediction.

The single-domain simulation, although computationally efficient (time, cost), overestimates summer precipitation, both in terms of amount and number of rainy days, and fails to capture drought events in croplands. In particular, this protocol produces more summer convective rain, associated with a higher summer cloud fraction than for the two-domain downscaling, particularly on low clouds. The two-domain downscaling performs better, accurately reproducing annual cycles, precipitation variability and the extreme 2003 drought, although it struggles with the less severe 2011 event. However, the two-domain downscaling amplifies positive biases in Tmax and PET, possibly due to overestimation of incoming shortwave radiation passing through reduced cloud cover and no nudging in the second domain. Bias correction for these variables may be necessary to avoid accelerated crop growth in impact models.

The performance of a direct downscaling of reanalyses to reproduce the local climate at less than 12 km (0.11°) over the Euro-CORDEX domain is questionable for territorial studies in Europe. Despite the limitations on Tmax and PET, the two-domain downscaling is a credible approach for agricultural studies and provides a reliable basis for analysing precipitation extremes and their impact on crops.