In the context of escalating climate change, accurate forecasting of sea surface temperatures (SSTs), SST anomalies (SSTAs), and marine heatwaves (MHWs) is crucial due to their profound impacts on marine ecosystems, biodiversity, and human societies. The increasing frequency and intensity of these oceanic temperature oscillations and extreme events necessitate advanced methodologies for reliable predictions to help mitigate their adverse effects. This thesis introduces MoanaNet, an integrated deep learning (DL) framework that enhances forecasting capabilities for SSTs, SSTAs, and MHWs by leveraging spatiotemporal anomaly forecasting methodologies, data science techniques, and domain knowledge.

The core innovation of this research lies in its comprehensive exploration of machine learning methods from anomaly, spatial, and temporal perspectives to advance data-driven forecasts of SSTs, SSTAs, and MHWs. By integrating these approaches into a cohesive framework, MoanaNet addresses the complexities inherent in predicting marine climate phenomena. This thesis evaluates the predictive performance of MoanaNet through spatial and temporal analyses, utilizing case studies to demonstrate areas where predictions are more or less successful. Comparisons with existing models and benchmarks further highlight the potential of MoanaNet to significantly improve forecasting of not only ocean temperature oscillations and extreme events but also broader climate phenomena.

The findings of this research contribute to ongoing efforts to develop robust data-driven predictive tools for environmental applications, which are substantially different from conventional physics-based numerical approaches. The proposed tools can aid in planning and executing marine conservation strategies and managing the effects of climate anomalies on human activities.

The specifi c contributions of the thesis can be summarized as follows:

A DL framework that combines three methodologies: imbalanced regression (Chapter 2), graph representation (Chapter 3), and temporal diffusion (Chapter 4), to enhance the forecasting of SSTAs and MHWs from short to long lead times on a global scale is proposed. To the best of my knowledge, this is the rst research to synergistically combine these methodologies to predict SSTAs and MHWs. Furthermore, the framework allows for potential application to other climate extreme forecasts. Each component's primary contribution is detailed below:

In Chapters 2 and 4, imbalanced regression and optimization of loss functions tailored for MHW prediction are investigated. Imbalanced regression loss functions, whether chosen from the existing or customized, are found to increase the detection of MHW occurrences, while the standard mean squared error (MSE) loss tends to lower the false alarm rate. Compared to the standard regression loss functions, this approach achieves higher average critical success index (CSI) or symmetric extremal dependence index (SEDI) scores. At 12 selected MHW hotspots, the imbalanced regression loss functions generally outperformed the standard MSE.

In Chapters 3 and 4, two new graph construction approaches designed to convert SST grids into SSTA graphs and generate new publicly available SSTA graph datasets are presented. Several classes of graph neural networks, focusing on the use of the GraphSAGE architecture, are investigated. Generally, the graph approach outperformed the persistence model in short-term global SST and SSTA forecasts, provided up-to-two-year-ahead SST forecasts with a recursive model, and improved MHW predictions both on average and at 12 selected MHW hotspot locations.

In Chapters 3 and 4, time series forecasting approaches for long-term prediction are integrated. Besides further examining the conventional sliding window approach and recursive method for climate forecasts, and the long short-term memory mechanism, an adaptation of the temporal diffusion process that eliminates the need for sliding windows in climate forecasting is introduced. The diffusion method is found to reduce input requirements and enhance long-term MHW predictive skills. To the best of my knowledge, this is the first attempt to combine temporal diffusion with graph neural networks for climate forecasts.

The outcomes of MoanaNet can be split in two categories: spatial and temporal. Spatially, in terms of one-month-ahead forecasts, the SST and SSTA models, evaluated by the root mean squared error (RMSE), performed better than the persistence model across most locations around the world. The MHW models, evaluated by the SEDI, generated forecasts with distinct spatial patterns compared to numerical models on a global scale.

Temporally, evaluated by the RMSE, the SST models produced forecasts up to two years ahead, with average RMSEs around 0.8 across all locations. Evaluated by combining the RMSE, precision, recall, CSI, and SEDI, SSTAs and MHWs could be predicted up to six months in advance.

In summary, this thesis contributes to machine learning for climate forecasting by introducing MoanaNet, a deep learning framework that improves the prediction of SSTs, SSTAs, and MHWs. By integrating imbalanced regression, graph representation, and temporal diffusion, MoanaNet aims to enhance predictive performance across spatial and temporal scales. The framework not only introduces new tools for climate forecasting but also establishes a foundation for future research in data-driven Earth system modeling, supporting efforts to manage marine ecosystems and mitigate the impacts of climate extremes.