Accurate Sea Level Anomaly Forecast over North Pacific with Improved AI Training Strategies

Timely and accurate sea level anomaly (SLA) forecasts are essential for monitoring the ocean environment and guiding coastal management. Utilizing multi-satellite altimetry products is effective for meeting these forecast demands. Here, we report an artificial intelligence-based SLA forecast model tailored for the North Pacific Ocean using altimetry products and a previously applied transformer-based network (i.e., Earthformer) to achieve short-range (<7 d) and medium-range (7 to 30 d) forecasts. Building upon the high baseline performance, when optimized training strategies are considered, medium-range prediction accuracy can be markedly improved. The 2 innovative training strategies include changing the prediction target to the SLA temporal tendency (to address the slow variation of the SLA) and bridging the training-forecast gap (training only for the next day, while forecasting for longer horizons) through rolling and multi-step training. Consequently, the formally trained model Multistep-Earthformer surpassed the persistence forecast (21% to 58% root mean square error reduction in the test over 1 to 30 prediction lead days) and state-of-the-art numerical product GLO12v4 (10% to 77% root mean square error reduction in the forecast over 1 to 10 prediction lead days). The results highlight the strong capability of Multistep-Earthformer in delivering accurate short- to medium-range SLA forecasts. Moreover, the 2 training strategies employed are model-agnostic and hold potential for enhancing various geoscientific prediction tasks.