Forecasting monthly-to-seasonal sea surface temperatures and marine heatwaves with graph neural networks and diffusion methods

Knowing future sea surface temperature (SST) patterns play a crucial role not only in industries such as fisheries, shipping and tourism but also in conservation of marine species . For example, DNA of endangered species can be sampled prior to anticipated marine heatwaves to preserve marine biodiversity. Overall, availability of SST forecasts allows to mitigate potential adverse impacts of extreme events such as marine heatwaves. 

There is a strong interest in accurate forecasts of SST and their anomalies on various time scales. The commonly used approaches include physics-based models and machine learning (ML) methods. The first approach is computationally intensive and limited to shorter time scales. While several attempts have been made by the community to adapt ML models to SST forecasts several challenges still remain. These challenges include improving accuracy for longer lead SST anomaly forecasts. 

Here we present an integrated deep-learning based approach to the problem of SST anomalies and MHW forecasting. On one hand, we capitalise both on inherent climate data structure and recent advances in the field of geometric deep learning. We base our approach on a flexible architecture of graph neural networks, well suited for representing teleconnections. From another hand, we adapt the diffusion method to increase lead time of the forecasts.  Our integrated approach allows marine heatwave forecasts up to six months in advance.