6 Conclusions

6.1 Summary of Research Findings

In Chapter 2, I examined the spatiotemporal variability and long-term trends in phytoplankton biomass using 26 years (1998–2023) of satellite-derived ocean-colour data. The analysis focused on seasonal, interannual, and decadal variability across major Southern Ocean frontal zones. Pronounced seasonal variability was identified, with summer chlorophyll-a (Chl-a) concentrations nearly twice those observed in winter. This pattern is consistent with previous studies linking Southern Ocean seasonality to variations in light availability, mixed-layer dynamics, and nutrient supply, although these mechanisms were not explicitly investigated in this thesis. . Interannually, peak Chl-a occurred in 2010, while the lowest values were observed in 2018, revealing a quasi-periodic variability of approximately 5–6 years. This finding indicates that Southern Ocean phytoplankton variability is influenced not only by long-term trends but also by substantial interannual fluctuations, highlighting the importance of considering multiple temporal scales when assessing ecosystem change. Spatial trend analyses indicated increasing Chl-a over approximately 60% of the Southern Ocean, particularly in coastal regions and along southern boundary zones, suggesting a basin-scale “greening” trend. This increase is likely associated with enhanced nutrient supply from upwelling and shelf-related processes. To identify the physical drivers of long-term Chl-a variability, regression analyses were conducted using sea surface temperature (SST), mixed-layer depth (MLD), photosynthetically active radiation (PAR), and wind speed (WS). SST emerged as the dominant driver of Chl-a anomalies across roughly 75% of the Southern Ocean, exerting a stronger influence than either MLD or WS. However, regional complexity was evident, with wind-driven mixing acting as a secondary control in frontal regions. These results also highlight the role of southward frontal migration in modifying phytoplankton habitats and community structure.

In Chapter 3, the analysis was extended to incorporate the influence of sea ice as a key physical modulator of phytoplankton variability. I examined the effects of sea ice extent anomalies on interannual Chl-a variability across major frontal zones, accounting for regime shifts and regional heterogeneity. The results revealed that summer sea ice extent has declined at a rate approximately five times faster than winter extent, leading to post-2015 regime shifts characterised by concurrent reductions in both sea ice and Chl-a, particularly in coastal and marginal ice zones. Positive sea ice anomalies were associated with elevated Chl-a concentrations in the Southern Boundary and ACC regions. Previous studies have linked such responses to changes in upper-ocean stability and nutrient availability, although these processes were not directly measured in the present study . These findings demonstrate that sea ice is a critical regulator of phytoplankton dynamics, with melt events triggering blooms in Antarctic coastal and frontal regions.

To assess the potential future evolution of Southern Ocean phytoplankton, Chapter 4 focused on reconstructing Chl-a distributions using a deep-learning U-Net architecture trained on satellite observations and reanalysis datasets. The model demonstrated strong reconstruction skill when evaluated against satellite Chl-a, successfully capturing bloom-prone regions and spatial variability patterns. Feature-importance analyses identified SST and sea level anomaly (SLA) as the primary predictors of phytoplankton variability. Model performance was highest in the Pacific sector, while reduced skill in coastal regions reflected data limitations and the influence of episodic processes. Building on this validated framework, the trained U-Net was applied to project future phytoplankton trajectories under CMIP6 climate scenarios, as described in Chapter 5. The projections revealed widespread declines in primary productivity that scale with greenhouse gas emission intensity, with more than 90% of the Southern Ocean exhibiting negative trends under the high-emission SSP5-8.5 scenario. Declines in Chl-a were associated with increasing SST and shoaling MLD across much of the Southern Ocean. These patterns may reflect changes in upper-ocean stability and a reduction in the depth of the surface reservoir, conditions that could limit nutrient entrainment from deeper waters and thereby constrain phytoplankton biomass. However, these mechanisms were not directly evaluated in this thesis and are inferred from established understanding of Southern Ocean physical-biological interactions.

By integrating the findings from Chapters 2–5, this thesis demonstrates the high sensitivity of Southern Ocean phytoplankton to physical and climatic forcing, with far-reaching implications for global biogeochemical cycles and climate regulation. While the observed increases in Chl-a suggest an enhanced potential for carbon sequestration and oceanic CO₂ uptake, projected declines under high-emission scenarios indicate a potential weakening of the biological carbon pump, thereby amplifying climate feedbacks. Ecologically, these changes may disrupt marine food webs and affect higher trophic levels. Regionally, the Pacific sector emerges as a critical hotspot requiring sustained monitoring, while ongoing sea ice loss underscores the need for integrated observational and modelling approaches. Overall, this research highlights the urgency of climate mitigation to safeguard Southern Ocean productivity and marine biodiversity.

6.2 Critical Reflection on Methodological Assumptions

Several methodological assumptions underpin the analyses presented throughout this thesis and should be considered when interpreting the results. A key assumption in Chapters 2 and 3 was the use of climatological frontal boundaries to investigate long-term chlorophyll-a variability across Southern Ocean frontal zones. This approach provided a consistent spatial framework for comparing regions over time; however, Southern Ocean fronts are dynamic features that shift in response to mesoscale circulation, wind forcing, and large-scale climate variability. Consequently, some of the trends identified within fixed frontal zones may partly reflect changes in frontal position rather than changes occurring entirely within stationary oceanographic regimes. Future analyses should therefore test the sensitivity of trend estimates to dynamically varying frontal positions derived from satellite altimetry or other frontal detection products.

Other methodological choices may also influence the magnitude and spatial expression of the reported trends. These include the selection of the GlobColour chlorophyll-a product, the definition of seasons, and the restriction of some validation analyses to 2017. Such choices provide necessary consistency for the analyses but may introduce uncertainties associated with sensor continuity, data availability, and the representation of seasonal variability. The reported patterns should therefore be interpreted in the context of these methodological constraints rather than as entirely independent of the underlying observational framework.

A further assumption concerns the interpretation of chlorophyll-a as an indicator of phytoplankton biomass and ecosystem variability. Although chlorophyll-a is widely used as a satellite-derived proxy for phytoplankton biomass, it is also strongly influenced by photoacclimation, pigment packaging, physiological state, and community composition. Changes in chlorophyll-a therefore do not necessarily correspond proportionally to changes in phytoplankton biomass or primary production. This distinction is particularly important in the Southern Ocean, where light availability, nutrient limitation, and community composition can vary substantially across seasons and frontal regimes. Accordingly, mechanisms discussed throughout this thesis—including changes in stratification, nutrient supply, phenology, and community composition—should be regarded as plausible interpretations supported by the observed environmental context and previous literature, rather than as processes directly demonstrated by the analyses.

The deep-learning framework developed in Chapters 4 and 5 introduces an additional set of assumptions. The U-Net framework was designed to reconstruct broad-scale chlorophyll-a variability from physical environmental predictors; however, several important ecological and biogeochemical controls were not explicitly represented. In particular, iron availability, grazing pressure, species-specific physiological responses, and biological interactions were excluded from the predictor set. This limitation may be particularly important in regions where iron limitation, shelf influence, sea-ice melt, sediment inputs, or lateral transport strongly regulate phytoplankton productivity. Consequently, the model should be interpreted primarily as a data-driven framework for identifying and reconstructing relationships between chlorophyll-a and the selected environmental variables, rather than as a complete mechanistic representation of Southern Ocean phytoplankton dynamics.

The projection framework in Chapter 5 further assumes that relationships learned from historical observations and historical CMIP6 simulations remain sufficiently stable under future climate conditions. This stationarity assumption becomes increasingly uncertain as environmental conditions move beyond the range represented in the training data, particularly under high-emission scenarios such as SSP5-8.5. Future changes in sea-surface temperature, sea-ice cover, mixed-layer dynamics, and upper-ocean stability may produce environmental states that are poorly represented in the historical training period. Under these conditions, the model may be required to extrapolate into novel environmental regimes, increasing uncertainty in late-century projections. The projected changes in chlorophyll-a should therefore be interpreted as conditional projections based on the persistence of the learned physical–biological relationships, rather than deterministic predictions of future ecosystem behaviour.

These assumptions collectively highlight an important distinction between statistical association and mechanistic attribution. The results provide evidence for relationships between physical environmental variability and chlorophyll-a changes across the Southern Ocean, but they do not fully resolve the underlying biological and biogeochemical mechanisms. This distinction is particularly relevant when interpreting model-derived relationships under future climate conditions.

6.3 Limitations and Future Research Directions

Beyond the methodological assumptions discussed above, several broader observational, biogeochemical, and modelling limitations constrain the current analysis. The reliance on surface chlorophyll-a may underestimate subsurface chlorophyll maxima and under-ice blooms, particularly in polar regions where light and mixing conditions vary seasonally. Persistent data gaps during high-latitude winters and in ice-covered regions further limit observational constraints, while uncertainties inherent in CMIP6 Earth System Models, especially in the representation of iron cycling and other micronutrient dynamics, introduce additional uncertainty into future projections. Moreover, the coarse spatial resolution in many models restricts the accurate representation of mesoscale processes, which are known to strongly influence nutrient supply and phytoplankton productivity.

Future research should prioritise the integration of in situ biogeochemical observations, such as those provided by biogeochemical Argo floats, gliders, and moored platforms, to improve validation of satellite-based reconstructions and to better constrain nutrient and micronutrient controls on phytoplankton growth. Incorporating additional environmental drivers—including iron availability, enhanced solar radiation metrics beyond photosynthetically active radiation, and surface circulation fields—would allow a more complete representation of coupled physical–biological processes. Increasing spatial resolution through downscaling or the use of eddy-resolving simulations (≤0.5°) would further improve the representation of frontal and mesoscale dynamics. From a modelling perspective, extending deep-learning frameworks to multi-output architectures capable of predicting phytoplankton functional types, using products such as GlobColour PFTs, would enable more ecologically meaningful assessments of community composition and carbon export. The application of explainability and uncertainty quantification methods, including SHAP-based attribution and ensemble modelling approaches, would enhance interpretability and confidence in future projections.

Finally, sustained long-term observations remain essential for evaluating whether the relationships identified in this thesis persist under continued climate change. Continued development of the Southern Ocean Observing System (SOOS), together with expanded observations in underrepresented sectors such as the Indian Ocean sector, would improve detection of emerging ecosystem regime shifts and provide essential constraints for future satellite and machine-learning products. Integrating sustained observations, improved biogeochemical measurements, higher-resolution modelling, and uncertainty-aware machine-learning approaches will ultimately be necessary to move from identifying statistical relationships toward a more mechanistic understanding of how Southern Ocean phytoplankton communities respond to climate variability and change.