Neglected Spatiotemporal Variations of Model Biases in Ensemble‐Based Climate Projections

Abstract The Bayesian model averaging (BMA) method has been widely used for generating probabilistic climate projections. However, the averaging weights used in BMA can only reflect the spatially‐ and temporally‐averaged performance of each ensemble member, without the ability to address the spatiotemporal variations of model biases. This can lead to inevitable exaggeration or understatement of the contributions of individual members to the ensemble mean, thus reducing the robustness of the resulting probabilistic projections. Here we propose a new method to help address the neglected spatiotemporal variations of model biases. Through the proposed method, the BMA weights are used as prior distributions to drive the Bayesian discriminant analysis in order to generate refined weights for individual ensemble models according to their spatially‐ and temporally‐clustered performance. Through applying the proposed method to Canada, we demonstrate its effectiveness in generating robust probabilistic climate projections (e.g., the average R 2 increases from 0.82 to 0.89). Plain Language Summary Since no climate model is superior to others under all conditions, taking the advantages of multiple models is desired for robust climate projections. However, there exist significant spatial and temporal variations regarding the model performances. For example, an overall poorly‐performed model can still show promising results over a specific region or during a specific time period, while an overall well‐performed model does not necessarily perform well over all regions or in all years. Here we propose a new method to robustly address the spatiotemporal variations in the modeling performance of each ensemble member. Then we apply the proposed method to generate the probabilistic temperature projections over Canada. The proposed method can improve both the accuracy and reliability in ensemble‐based climate projections. Key Points A discriminant‐BMA ensemble modeling approach is developed to synthesize the spatiotemporal variations of model biases The accuracy and reliability of probabilistic projections are both improved (e.g., the average R 2 increases from 0.82 to 0.89) The temperature probabilistic projections based on regional climate models in three future periods are generated over Canada