Improved short-term sea level change predictions with better AI training

(Press-News.org) Sea level can temporarily change for a variety of reasons—atmospheric pressure shifts and water accumulation from wind and storms, for example—which can cause flooding in coastal communities and affect maritime industry operations. The key to mitigating the effects of short-term sea level variation is accurate prediction that provides ample warning time to affected areas.

 

Sea level anomaly (SLA) is a key parameter for predicting short-term variations in sea level. It reflects the absolute geostrophic current anomaly, which measures the difference between the near-term ocean surface currents and the long-term average. SLA is derived from satellite altimetry observations, which provide precise measurements of sea surface height variations on a global scale.

Numerical models are the primary tool used for short-term SLA predictions. Incorporating the information from satellite altimetry observation, numerical models can provide accurate and timely forecast of SLA. Despite this, today’s numerical models still contain persistent biases and excessively high computational costs, which have prompted researchers to explore new ways to model and predict short- and medium-term changes in sea level.

AI has demonstrated great potential in data-driven prediction across a wide range of marine systems. Large-scale AI models have outperformed other state-of-the-art numerical models over 10-day forecasts. These Global Ocean Forecast Systems (GOFSs), however, are designed to work on a global scale, and also require excessive computing power for regional applications. Instead, researchers require an SLA forecasting model that combines regionally optimized training protocols with observational data.

To address this issue, researchers from Sun Yat-Sen University, Zhejiang Institute of Marine Planning and Design, and Pusan National University focused on enhancing sea level prediction accuracy in the North Pacific Ocean without increasing the architectural complexity of AI models.

 

The team published the study on January 23 in the journal Ocean-Land-Atmosphere Research (OLAR).

“[Our goal was] to achieve high-accuracy short- and medium-range sea level anomaly forecasting by optimizing training strategies, rather than simply making the AI models more complex,” said Jiangnan He, a graduate student in the School of Marine Sciences at Sun Yat-Sen University in Zhuhai, China and first author of the research paper.

 

By focusing on improved AI model training, the team mitigated error accumulation and limitations inherent in all models when integrating predictions over time. Additionally, the researchers decided to target daily variation in SLA data, rather than absolute values, in an attempt to improve forecast accuracy.

 

“We developed two training strategies and used them to build a new forecasting model, which delivers much more accurate predictions than existing approaches. Moreover, these strategies can also be widely applied to other AI-based prediction tasks,” said Wenfang Lu, an associate professor in the School of Marine Sciences at Sun Yat-Sen University and corresponding author of the research paper.

 

The new AI-based SLA forecasting model is based on the Earthformer, a state-of-the-art AI model, and tailored for the North Pacific Ocean using altimetry data and a deep-learning model capable of processing information in parallel, rather than sequentially. From there, the researchers optimized two training strategies to greatly improve medium-range prediction accuracy. First, the team changed the prediction target to the SLA temporal tendency to reflect the slow, long-term variation of the SLA. Secondly, the team addressed the training-forecast gap, training the model only for the next day while forecasting for longer horizons through rolling and multi-step training.

Overall, the team’s trained SLA forecasting Multistep-Earthformer model outperformed persistence forecasts, a naïve forecast method assuming future conditions will mimic current or past observations, and the current state-of-the-art numerical GLO12v4 model. Additionally, the strategies used to train the Multistep-Earthformer model can be used with any AI model, and could further enhance other geoscientific prediction tasks.

 

Given the team’s success in modeling the North Pacific Ocean, they are looking to extend the model’s application and training optimization to other regions and disciplines. “Our next step is to expand the study area to the global ocean and develop training strategies tailored to different ocean regions. Our ultimate goal is to further optimize these strategies and build a reliable AI-based global sea level forecasting system to better support ocean monitoring and practical applications,” said He.

 

Yong Liu from the Zhejiang Institute of Hydraulics & Estuary (Zhejiang Institute of Marine Planning and Design) in Hangzhou, China; Guangyu Yang from the School of Marine Sciences at Sun Yat-Sen University in Zhuhai, China; Young-Heon Jo from the Department of Oceanography and Marine Research Institute at Pusan National University in Busan, Republic of Korea; and Zhigang Lai from the School of Marine Sciences at Sun Yat-Sen University in Zhuhai, China, the Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai) in Zhuhai, China, and the State Key Laboratory of Environmental Adaptability for Industrial Products at Sun Yat-Sen University in Guangzhou, China also contributed to this research.

 

This study was supported by Guangdong Basic and Applied Basic Research Foundation: 2025A1515012024 and Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai): SML2023SP238.

END