Deep Learning Weather Models for Subregional Ocean Forecasting: A Case Study on the Canary Current Upwelling System
Giovanny Alejandro C-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA), ECOAQUA Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Javier Sánchez Centro de Tecnologías de la Imagen (CTIM), IUCES Universidad de Las Palmas de Gran Canaria Las Palmas, Spa · 2025
Read Deep Learning Weather Models for Subregional Ocean Forecasting: A Case Study on the Canary Current Upwelling System by Giovanny Alejandro C-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA), ECOAQUA Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Javier Sánchez Centro de Tecnologías de la Imagen (CTIM), IUCES Universidad de Las Palmas de Gran Canaria Las Palmas, Spa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-24429.txt.
Introduction
Ocean forecasting for a rich upwelling zone is the aim of a 2025 study by Cuervo-Londoño and co-authors. The Canary Current system feeds major fisheries, but classic physics models are slow and costly to run. The team adapted a graph neural network, first built for global weather, to this smaller sea region.
Trained on satellite data, the model was checked against top physics models and a reanalysis product. It cut forecast errors most in spots with complex currents, such as three named capes. Five-day forecasts at those capes improved by up to three quarters against the reanalysis. Some trouble spots remained in upwelling areas, the authors note.
Catalog record
| CREATOR | Giovanny Alejandro C-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA), ECOAQUA Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Javier Sánchez Centro de Tecnologías de la Imagen (CTIM), IUCES Universidad de Las Palmas de Gran Canaria Las Palmas, Spa |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 93 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-24429
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Deep Learning Weather Models for Subregional Ocean Forecasting: A Case Study on the Canary Current Upwelling System Giovanny Alejandro C-Londoño Oceanografía Física y Geofísica Aplicada (OFYGA), ECOAQUA Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Javier Sánchez Centro de Tecnologías de la Imagen (CTIM), IUCES Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Ángel Rodríguez-Santana Oceanografía Física y Geofísica Aplicada (OFYGA), ECOAQUA Universidad de Las Palmas de Gran Canaria Las Palmas, Spain [email protected] Abstract Oceanographic forecasting impacts various sectors of society by supporting environmental conservation and economic activities. Based on global circulation models, traditional forecasting methods are computationally expensive and slow, limiting their ability to provide rapid forecasts. Recent advances in deep learning offer faster and more accurate predictions, although these data-driven models are often trained with global data from numerical simulations, which may not reflect reality. The emergence of such models presents great potential for improving ocean prediction at a subregional domain. However, their ability to predict fine-scale ocean processes, like mesoscale structures, remains largely unknown. This work aims to adapt a graph neural network initially developed for global weather forecasting to improve subregional ocean prediction, specifically focusing on the Canary Current upwelling system. The model is trained with satellite data and compared to state-of-the-art physical ocean models to assess its performance in capturing ocean dynamics. Our results show that the deep learning model surpasses traditional methods in precision despite some challenges in upwelling areas. It demonstrated superior performance in reducing RMSE errors compared to ConvLSTM and the GLORYS reanalysis, particularly in regions with complex oceanic dynamics such as Cape Ghir, Cape Bojador, and Cape Blanc. The model achieved improvements of up to 26.5% relative to ConvLSTM and error reductions of up to 76% in 5-day forecasts compared to the GLORYS reanalysis at these critical locations, highlighting its enhanced capability to capture spatial variability and improve predictive accuracy in complex areas. These findings suggest the viability of adapting meteorological data-driven models for improving subregional medium-term ocean forecasting. This also demonstrates the superior flexibility of graph neural networks compared to traditional models, as they can be adapted to new prediction tasks even when originally developed for different purposes. K eywords Sea surface temperature forecasting; Graph neural networks; Canary Current Upwelling System; Data-driven ocean prediction; Operational oceanography 1 Introduction Oceanographic prediction is crucial for understanding climate change and supporting sectors like maritime transport, fisheries, and natural disaster management ( Bell et al., 2009 ) . It relies heavily on accurately forecasting mesoscale processes due to their environmental and economic impacts. These processes give rise to distinct structures, influencing mean currents and transporting key ocean properties ( Falkowski et al., 1991 ) . Despite its importance, predicting mesoscale processes remains challenging for operational forecasts ( Treguier et al., 2017 ; Mourre et al., 2018 ) , as evidenced by the difficulty in forecasting the Gulf of Mexico Loop Current eddy during the 2010 Deepwater Horizon oil spill ( Adcroft et al., 2010 ; Liu et al., 2013 ) . Traditional oceanographic prediction techniques rely on numerical models that solve physics-based equations. While these models have significantly advanced ocean forecasting, the theoretical foundation—quasi-geostrophic (QG) theory—and the numerical models have inherent limitations that hinder their accuracy in certain contexts. Although the QG theory initially advanced our understanding of mesoscale dynamics and ocean prediction, it struggles to address strong currents, cannot account for bathymetric features, fails to incorporate surface density gradients, and does not model frontal dynamics ( Cushman-Roisin et al., 1990 ) . Despite these theoretical limitations, global ocean circulation models remain the primary tools for oceanographic forecasting, although they contribute to the persistent challenges of operational mesoscale prediction. On the other hand, numerical ocean prediction (NOP) models—such as NEMO ( Madec et al., 2024 ) , which forms the core of reanalysis systems like GLORYS ( Jean-Michel et al., 2021 ) and operational forecast systems like PSY4V3R1 ( Lellouche et al., 2018 ) —remain the current standard for short-term deterministic forecasting. Nevertheless, they cannot accurately represent reality due to diverse constraints such as incomplete understanding of subgrid-scale parameterizations, poorly known forcing fields, insufficient knowledge of interactions with other Earth system components, and restricted computational resources ( Sommer et al., 2018 ) . Additionally, incomplete observations with spatiotemporal gaps and the limitations of data assimilation schemes—still in continuous improvement—prevent these models from fully capturing the ocean’s state. Furthermore, these models do not fully utilize extensive historical data and lack optimization for modern hardware, such as GPUs, which further reduces their efficiency. These limitations highlight the need for novel approaches to improve ocean prediction capabilities. In recent years, short-term machine learning weather prediction (MLWP) models have emerged in global atmospheric forecasting, surpassing the efficiency and accuracy of traditional numerical systems ( Bouallègue et al., 2024 ) . Models such as Pangu Weather ( Bi et al., 2023 ) , GraphCast ( Lam et al., 2023 ) , Aurora ( Bodnar et al., 2024 ) , NeuralGCM ( Kochkov et al., 2023 ) , Gencast ( Price et al., 2024 ) , or AIFS ( Lang et al., 2024 ) have demonstrated the power of this approach, significantly reducing inference times and computational costs. By leveraging historical data and focusing on spatiotemporal patterns, MLWP models bypass the constraints of incomplete physical understanding and adapt easily to new prediction tasks without architectural modifications ( Dueben and Bauer, 2018 ; Scher, 2018 ; Bi et al., 2023 ) . Nevertheless, MLWP models heavily depend on data quality and availability, facing challenges with heterogeneous, sparse, spatially discontinuous, or noisy datasets, which can limit performance in certain contexts. Additionally, the physical consistency of these models deteriorates over long timescales due to spectral bias, which can result in numerical instabilities or unrealistic hallucinations ( Chattopadhyay et al., 2023 ) . While MLWP models perform well in atmospheric systems, their application to oceanography presents distinct challenges. Unlike atmospheric data, which is abundant, spatially continuous, and less influenced by external factors, oceanographic data is constrained by atmospheric interference, which affects observational accuracy and consistency, spatial discontinuities, predominantly induced by the presence of continental landmasses, which significantly disrupt data continuity and turn the training of these models complex. Recent advancements in machine learning ocean prediction (MLOP) have led to the development of innovative models at global and regional scales. Xihe ( Wang et al., 2024 ) , for instance, has been introduced for global ocean forecasting, while SeaCast ( Holmberg et al., 2024 ) and OceanNet ( Chattopadhyay et al., 2024 ) focus on regional applications. Similarly, progress in mesoscale ocean forecasting, including studies on eddy shedding predictability in the Gulf of Mexico ( Zeng et al., 2015 ) and data-driven turbulence forecasting using autoregressive techniques ( Chattopadhyay et al., 2021 ) , has shown promising results. The Canary current upwelling system (CCUS), the only eastern boundary upwelling system with islands, presents a unique prediction challenge. In this region, intense mesoscale activity results from the interplay between oceanic features, coastal regimes, and bathymetry, driving strong mesoscale stirring and creating a highly dynamic environment ( Sangrà et al., 2009 ; Arístegui et al., 1994 ; Barton et al., 1998 ) . These dynamics create a complex environment for sub-regional