Resumen de Near-term climate change: High-resolution decadal climate predictions in the Iberian Peninsula
Juan José Rosa Cánovas
The time scales of climate assessments play a crucial role in the development of climate services, which aim at effectively transforming the results of the scientific research into solutions to real-world problems. While the seasonal-to-interannual climate predictions take advance of an accurate description of the initial climate state to carry out estimations of the actual climate evolution in the near future, the long-term climate projections represent the potential climate evolution as a response to anthropogenic external forcings, in scales ranging from several decades to centuries, along different future scenarios of climate change. The decadal climate predictions (DCPs), the subject of study in this Thesis, bridge the gap between seasonal-to-interannual predictions and climate projections. At the decadal time scale, both initial conditions and external factors jointly contribute to the estimation of the climate signal, making DCPs valuable sources of climate information for a wide range of users and decision-makers in the social, environmental and economic spheres.
The main purpose of this Thesis has been to generate a collection of high-resolution DCPs over the Iberian Peninsula (IP) and evaluate their accuracy and reliability along with their added value over a global decadal prediction system (DPS) and a set of high-resolution uninitialized experiments. To address this task, a set of dynamical downscaling (DD) simulations was conducted with the Weather Research and Forecasting model (WRF), with data from the global CESM Decadal Prediction Large Ensemble (CESM-DPLE) as input information for the decadal experiments, from the global CESM Large Ensemble (CESM-LE) for the uninitialized simulations and from the ERA-Interim reanalysis for other additional experiments. The DD simulations were carried out in two nested domains. A coarse-grid domain was defined to cover the EURO-CORDEX region with an horizontal resolution about 50 km, whereas a fine-grid domain, with an approximate resolution of 10 km, was centered in the IP. Approximately 4.94 million CPU hours were dedicated to conduct the DD simulations required to produce a total of 1470 simulated years. To the best of my knowledge, the research presented here constitutes the first study which comprehensively assesses the performance of a dynamically downscaled DPS at an horizontal resolution of 10 km, becoming the maximum resolution attained in this branch of the climate prediction.
In spite of the huge development achieved in climate modeling during the last three decades, models are intrinsically based on approximations and, consequently, contain biases which arise from different sources. Thus, the CESM-DPLE and CESM-LE datasets were bias corrected before conducting the DD simulations to reduce the potentially negative impact that those biases may have on the downscaled product. The simulations were conducted in a context of limited access to computing resources, so a representative subset of 4 members for each global ensemble was selected to provide the input information for the simulations, since the task of downscaling the whole ensembles was not addressable. For each member of the CESM-DPLE subset, the experiments initialized every year from 1970 to 1999 were downscaled, as well as the whole 10-member ensemble available for DD along the decade initialized in 2015. For CESM-LE, however, the simulated period is shorter because the data were only available from 1990 to 2005.
The evaluation of the downscaled product has been focused on four primary climate variables: precipitation and maximum, minimum and mean near-surface air temperature. Before the analysis, the dynamically downscaled decadal experiments were recalibrated to reduce the unconditional and conditional biases and adjusting the ensemble spread of the WRF output fields. Significant improvements over the global CESM-DPLE and the downscaled uninitialized experiments have been found at both annual and seasonal scales for temperature variables, whereas the added value of the downscaled DCPs to the predictive skill for precipitation is more limited. The signal-to-noise paradox is strong in the predictions for precipitation and, to a lesser extent, also for temperature. The results suggest that high improvements in the predictive skill may be achieved by adding new members to the downscaled ensemble to compute larger ensemble averages and thus reduce the unpredictable background noise in predictions, especially for precipitation.
The sensitivity of WRF simulations to extreme initial conditions of soil moisture has also been examined. A set of simulations was conducted with ERA-Interim providing the input information for all variables with the exception of soil moisture. Three different types of soil moisture initial conditions were considered to represent a wet, a dry and a very dry soil. These initial conditions were calculated by combining the soil moisture index with some physical soil properties which depend on the soil textures. To account for the impact of the initialization date of the simulations, they started in two different dates, 1990-01-01 and 1990-07-01, covering in both cases the 10-year periods up to 1999-12-31 and 2000-06-30, respectively. The analysis of these simulations has been focused on the influence of the initial conditions on the spin-up requirements for soil moisture, precipitation and the three temperature variables. A maximum spin-up time of 8 years is needed for soil moisture in some cases, decreasing down to values generally lower than 3 years and 2 years for maximum and mean temperature, respectively. The spin-up time is commonly lower than 1 year for minimum temperature and mostly below 10 months for precipitation.
Since no spin-up time has been considered in the analyses of the downscaled decadal experiments initialized from 1970 to 1999 (it would have implied the loss of the first simulated years), the predictive skill might have experienced some deterioration because of the spin-up-related biases, at least during the first years of the simulations. Therefore, a dynamically equilibrated soil state, taken from a control WRF simulation, was used to initialize the simulations conducted for the decadal experiments in the decade 2015-2025 with the aim of improving as much as possible the predictive skill of the downscaled predictions. In regions with reliable predictions for precipitation at annual scale, the predicted anomalies for this variable are generally positive at the beginning of the decade and turn into negative during the second half, with the Pyrenees and the Central System among the areas with the strongest negative anomalies. The predictions for the temperature variables show positive anomalies throughout the entire decade over the whole domain at annual scale. The highest anomalies have been found in summer, with values up to 2 K at the end of the decade in some southeastern and northeastern locations of the IP.
Finally, a set of alternative correction methods has been examined to improve the bias correction in CESM-DPLE experiments and thus produce a more skilful downscaled product in potential future experiments. The correction of the trend performs well and contributes to producing robust predictions for the North Atlantic Oscillation, becoming a suitable method to correct the input data for DD simulations focused on Europe or North America. On the other hand, a method based on considering reference initial conditions in the correction algorithm generally get overall sligthly better results than the other analyzed methods for the prediction of the El Niño/Southern Oscillation. This method may be preferable for DD simulations targeting South America. The dependence of the predictive skill on the ensemble size has also been analyzed with data corrected with these methods. A modest 3-member ensemble has shown to be a good alternative to larger ensembles in a context of limited computing resources for some specific applications.
The research presented in this Thesis evidences the valuable role that WRF can play for the generation of high-resolution decadal experiments over the IP, demonstrating the ability of produce skilful predictions despite the limitations imposed by the restricted access to computing resources. The multiple applications of DD in the branch of the DCP and their potential ramifications open a vast field of research which could be explored in future works by taking the study presented here as a solid starting point.
