Resolving Marine Ecosystem Dynamics in Time and Space with Machine Learning Approaches

This thesis explores the application of advanced machine learning techniques to better understand and predict phytoplankton dynamics, which are important for maintaining marine ecosystems and regulating global biogeochemical cycles. As primary producers, phytoplankton are foundational to the marine food web and play a significant role in supporting marine life and contributing to global environmental processes. However, predicting their response to environmental variability remains a challenge due to the complexity of marine ecosystems and the interactions between physical, chemical, and biological processes. Traditional numerical models have been limited in capturing these non-linear dynamics, prompting the need for more sophisticated approaches. In this work, we have progressively adopted machine learning techniques, moving from traditional approaches to deep learning and probabilistic models, to address these limitations. The research in this thesis provides a framework for modelling phytoplankton dynamics by integrating various methods to enhance predictive accuracy and capture uncertainty.

The research begins by exploring the impacts of ocean extremes, specifically marine heatwaves (MHWs), on phytoplankton populations. MHWs are characterised by prolonged periods of abnormally high sea surface temperatures and have far-reaching consequences for marine ecosystems. These extremes are increasing in both frequency and intensity due to climate change. The impact of MHWs on phytoplankton concentrations is highly variable, where some regions are showing a marked decline while others exhibit minimal changes. This variability underscores the importance of identifying and understanding regional drivers, such as changes in ocean currents, nutrient upwelling, and eddy dynamics to accurately predict phytoplankton responses to climate-induced ocean changes. Therefore, this initial study serves as a foundational analysis of the influence of large-scale oceanographic anomalies on phytoplankton and demonstrates the need for more sophisticated modelling approaches to capture the complexity of these interactions.

In the subsequent research phase, the thesis explores the use of deep learning models to understand the seasonal and spatial drivers of phytoplankton variability. By integrating physical variables such as surface temperature, salinity, and ocean currents derived from satellite data, this work demonstrates how machine learning can enhance our understanding of phytoplankton behaviour. Traditional models often struggle with the complex interplay between these variables, especially when attempting to model large-scale patterns or predict over longer periods. Deep learning models seem to be better equipped to identify and learn such patterns, providing more accurate spatial and temporal predictions. This thesis section demonstrates how integrating physical oceanographic data with machine learning techniques can offer new insights into the relative importance of variables in understanding underlying dynamics, while also enhancing model generalisation across different regions.

The final phase takes this exploration a step further by introducing a probabilistic deep learning approach to address inherent uncertainty in ecological forecasting. While deterministic deep learning models provide high accuracy, they often lack interpretability and exhibit overconfidence in their predictions. However, the probabilistic framework offers an alternative approach that not only generates predictions but also quantifies the uncertainty associated with each prediction. This allows for the quantification of uncertainty in both the model and the data, which is particularly important in marine ecosystems, where data sparsity, variability, and measurement errors pose significant challenges. This phase focuses on forecasting chlorophyll-a concentrations, with the added benefit of providing uncertainty estimates alongside predictions. This probabilistic approach opens new avenues for enhancing the interpretability of the models, enabling more informed decisions in marine management and conservation efforts.

Together, the three phases of research present an integrated approach to address the challenges of modelling phytoplankton dynamics. Initially, focusing on the global ocean, followed by a shift to the Atlantic Ocean, and finally converging on the regional level in the North Atlantic Subpolar Gyre. An evolution in the modelling approach is reflected in the progression from classical machine learning in the first phase, to deep learning in the second, and ultimately to probabilistic Bayesian methods in the third, aiming to overcome the limitations of each method. While traditional models offer insights into large-scale patterns, they are often not able to deal with the inherent uncertainty in complex ecosystems. Deep learning models have advanced our understanding of seasonal drivers and improved predictive capabilities, however, they still fall short in accounting for uncertainty. Bayesian approach, with its probabilistic framework, offers a solution to this problem, allowing for more transparent and reliable forecasts.

In conclusion, this thesis highlights the critical need for innovative modelling techniques to better understand and predict marine ecosystems. By advancing beyond traditional methods and incorporating machine learning approaches, it provides a framework for more accurate and reliable predictions of phytoplankton dynamics, which are crucial for sustaining the health of the global ocean. The integration of data-driven models with uncertainty quantification has far-reaching implications, not only for advancing scientific understanding but also for informing policymakers and conservationists in addressing the impacts of climate change on marine environments.