Multi-model approach for statistical downscaling and bias correction of Earth system model outputs for marine ecosystem impact applications

Earth system models (ESMs) are the main tool for understanding the impacts of global change and are regularly updated to provide more reliable scenarios of the future. However, their confrontation with observations reveals biases that need to be corrected, especially for impact applications where the absolute scale of the environmental variable is relevant. In addition, marine regional impact studies require fine-scale projections for strategic planning and management actions. Statistical downscaling provides a fast way to produce regional forcings from ESMs and can additionally produce bias-corrected outputs necessary for marine impact applications driven by or fitted to observed data. Statistical downscaling can use different parametric distributions depending on the variables used, and generalised regression can provide a flexible approach for this purpose. We propose a multi-model approach based on non-parametric generalised regression and a set of indicators to select a robust statistical downscaling model that can be used to project future scenarios for marine ecosystems. The empirical cumulative distribution of the variables to be downscaled is modelled, ensuring that not only the mean but also the variance and quantiles (including minima and maxima) are properly represented, improving the prediction of extreme events and taking into account spatial autocorrelation. We incorporated future bias indicators alongside traditional evaluation metrics for model selection to identify and mitigate potential extrapolation errors in future scenario projections, ensuring more robust and plausible downscaled climate outputs. The approach presented here is applied to two contrasted regional case studies, the Bay of Biscay-Celtic Sea ecosystem and the Northern Peru Current ecosystem, using sea surface temperature from the IPSL-CM5A-LR ESM. The results show that a multi-model selection approach is appropriate, as individual model performance is case-specific.