Seasonal evolution prediction has become increasingly important for climate risk management, with profound impacts on human lives and socioeconomic activities. However, when the prediction target shifts from seasonal means to the spatiotemporal evolution of climate anomalies, the direct outputs of even the current best dynamical models exhibit limited predictive skill. This limitation highlights a fundamental issue of predictability that calls for a re-examination of prediction targets. Here, we propose a new perspective for enhancing predictability by focusing on spatiotemporal evolution patterns (STEPs) as new prediction targets. These STEPs possess coherent joint spatiotemporal structures and exhibit slowly varying, predictable characteristics, enabling more targeted identification of key influencing factors and physically meaningful precursors, and thereby improving seasonal evolution predictions. Our experimental results clearly validate such a new perspective through significantly improved seasonal evolution predictions of the East-Asian summer rainfall anomalies outperforming those directly predicted from the C3S multi-model ensemble mean. The proposed STEP-based framework demonstrates strong extensibility across regions, climate phenomena, physical mechanisms, and spatiotemporal scales. Moreover, it facilitates deeper integration of dynamical, statistical, and artificial intelligence-based approaches, offering a practical pathway toward advancing seamless evolution prediction in the future.

季节预测,时空演变