Organizations: University of California Davis, Davis, CA, 95616, USA · University of Washington, Seattle, WA, 98195, USA · NOAA Geophysical Fluid Dynamics Laboratory, Princeton, NJ, 08540, USA · Massachusetts Institute of Technology, Cambridge, MA, 02139, USA
Marine ecosystems are increasingly impacted by climate change, necessitating tools to identify and predict spatial habitat information. To build such tools, ecological marine provinces, "eco-provinces", ecologically meaningful regions in the global ocean can be used. We use unsupervised machine learning (ML) to identify eco-provinces with corresponding uncertainty measures based on output of a global simulation of phytoplankton functional types. Our work aims to create a proof of concept to predict eco-provinces based on satellite ocean color data. To do so, we develop a hierarchy of explainable dense ensemble networks to infer how well the eco-provinces can be detected from modeled ocean color fields. Key results include that the delineated eco-provinces are both ecologically meaningful and can be inferred with high skill. However, no straightforward relationship was found where adding more input data to the network consistently improves inference skill, and there is an intricate tradeoff between inputs and prediction fidelity. Our work is a case for optimism and a cautionary tale of needing uncertainty quantification and careful validation of prediction fidelity.
Figures & tables
Figure 1: NEMI PFT Inputs and DNE Inputs . Panel a) displays in the left column boreal summer (June, July, August: JJA) and in the right column boreal winter (December, January, February: DJF) NEMI PFT concentrations. The PFTs are Diatoms (Diat), Coccolithophores (Cocco), Mixotrophic Dinoflagellates (Mixo), Picoeukaryotes (Pico-e), and Picoprokaryotes (Pico-p). Panel b) presents boreal summer (June, July, August: JJA) and boreal winter (December, January, February: DJF) DNE inputs. Included is irradiance reflectance for wavebands 425 nm (blue reflectance), 550nm (green reflectance), chlorophyll (chl), and phytoplankton carbon biomass (phyto). The bottom right table indicates which inputs are included in each version of the DNE.
Figure 2: DNE Construction . The construction of the DNE, including the modeled input data that can be remotely sensed, a grid cell by grid cell input into the robust DNE prediction, repetition for all grid cells, average prediction and entropy across the ensemble, and SHAP to determine the importance of each input for correct and incorrect grid cell predictions is displayed.
Figure 3: Embedded and spatial Eco-Provinces With Corresponding Uncertainty, and PFT Input Composition . Section a) presents the 10 JJA boreal summer and DJF boreal winter eco-provinces plotted on the embedding, which is produced by UMAP dimension reduction. The corresponding entropy (uncertainty) is plotted in b). Sections c) and d) display the 10 JJA and DJF eco-provinces plotted geographically, with the corresponding uncertainty measured with entropy in e) and f). g) displays the phytoplankton functional type composition for each eco-province.
Figure 4: V1 predicted eco-provinces, with corresponding true-false and entropy plots . Sections a) and b) illustrate the V1 averaged prediction of 2005 boreal summer (June, July August: JJA) and boreal winter (December, January, February: DJF) eco-provinces. V1 includes blue and green reflectance as DNE inputs. Sections c) and d) present where the prediction was correct in green and incorrect in purple, compared to the actual NEMI eco-province labels. Note the incorrect predictions in the Arctic, Western Pacific, and in between eco-provinces. Sections e) and f) display the averaged entropy (uncertainty) of the prediction. The incorrect prediction of the boreal summer (JJA) Western Pacific has fairly low uncertainty.
Figure 5: V2 predicted eco-provinces, with corresponding true-false and entropy plots . Fig. 5 has the same format as Fig. 4 , but for V2 which has chlorophyll as a DNE input, in addition to the blue and green reflectance in V1. Boreal summer corresponds to June, July, August (JJA) and boreal winter corresponds to December, January, February (DJF). Compared to V1, the prediction improves in the North Atlantic, but gets worse in the Western Pacific and the Southern Ocean. Note the high entropy in the incorrect Southern Ocean prediction, but low entropy associated with the Western Pacific incorrect prediction.
Figure 6: V3 predicted eco-provinces, with corresponding true-false and entropy plots . Here are the boreal summer (June, July, August: JJA) and boreal winter (December, January, February: DJF) results of V3, which includes phytoplankton carbon biomass as an input in addition to the inputs in V2. Note the continued improvement in the Arctic, alongside sustained difficulty in the summer (JJA) Southern Ocean and Western Pacific. As previously observed, the uncertainty of the incorrect Southern Ocean prediction is high compared to the uncertainty of the incorrect Western Pacific prediction.
Figure 7: SHAP results for Eco-Provinces (EPs) 0, 5, and 7 for all three versions of the DNE, including blue and green reflectance inputs. For boreal summer (JJA) and boreal winter (DJF) the importance of blue and green reflectance decreases across versions with the addition of chlorophyll and phytoplankton carbon biomass. Blue reflectance is more important for prediction on the boundary of eco-province 0, compared to green reflectance which is more influential in the center of the eco-province. Also note the high accuracy of predictions for these core eco-provinces 0, 5, and 7, which in S3 correspond to fairly high DNE inputs. More specifically, eco-province 0 has high levels of all inputs, eco-province 5 has high levels of all inputs except for blue reflectance, and eco-province 7 has medium levels of blue reflectance and fairly high levels of the other inputs.
Figure 8: SHAP results for Eco-Provinces (EPs) 0, 5, and 7 for all three versions of the DNE, including chlorophyll and phytoplankton carbon biomass inputs. Displayed are the boreal summer (June, July, August: JJA) and boreal winter (December, January, February: DJF) SHAP values for eco-provinces 0, 5 and 7 for chlorophyll and phytoplankton carbon biomass inputs across all three versions. Overall, chlorophyll is most influential in these predictions. Note that chlorophyll is more important in some cases for predicting the inside region of eco-province 0, compared to phytoplankton which is more important on the outside of the eco-province.
Figure 9: SHAP values for eco-provinces (EPs) 3 and 6 for versions 2 and 3, including blue and green reflectance inputs . The importance of blue and green reflectance across all three versions for the boreal summer (June, July, August: JJA) and boreal winter (December, January, February: DJF) prediction of eco-provinces 3 and 6 is demonstrated. Both inputs contribute to the incorrect predictions of these eco-provinces. Note that the incorrect prediction of eco-province 3 in the boreal summer (JJA) Western Pacific is much more surprising compared to the incorrect prediction of eco-province 6 in the boreal summer (JJA) Southern Ocean, due to the low uncertainty of the eco-province 3 incorrect prediction.
Figure 10: SHAP results for eco-provinces (EPs) 3 and 6, for chlorophyll and phytoplankton carbon biomass as inputs, across versions 2 and 3 . The importance of chlorophyll to the incorrect prediction of boreal summer (June, July, August: JJA) eco-province 3, especially in V3 which includes all four inputs, is emphasized. For eco-province 6, chlorophyll is most influential to the incorrect prediction in V2, while phytoplankton carbon biomass is most influential for V3.
Figure S1: 20-Cluster Example. Example of an eco-province characterization with 20 eco-provinces (clusters) and the corresponding entropy percentage.
Figure S2: Training, validation, and test split. Fig. S2 demonstrates the vertical split used to train, validate, and test the neural network with input data averaging across 1993-2004. For an extra level of validation, we then tested the network on the entire globe for 2005.
Figure S3: DNE Input Composition . Fig. S3 presents the scaled composition of blue reflectance, green reflectance, chlorophyll, and phytoplankton carbon biomass as inputs to the V3 DNE, broken up according to each eco-province. This composition allows us to identify the input composition for each different spatial region, in order to understand which inputs contributed to specific results. In general, either high or low input concentrations are more informative for DNE predictions compared to moderate concentrations.
Table S1: Summer (JJA), Winter (DJF), and Average Accuracy Across Versions. Table S1 demonstrates the variety in accuracy among the three versions, revealing the increase in overall performance with the addition of inputs with some nuance in performance when considering each cluster individually.
Figure S4: Ensemble loss and accuracy comparison across three versions.
Figure S5: Ensemble loss and accuracy comparison across three versions, with minimum loss and maximum accuracy values. Fig. S5 also compares the ensemble losses and accuracies across the three versions, with minimum loss values and maximum accuracies.
Phytoplankton play a central role in marine ecosystems and the global carbon cycle, with different groups contributing differently to ocean biogeochemical processes. While standard techniques exist for monitoring phytoplankton concentration from ocean-colour data, their community composition remains difficult to observe at large scales. Chlorophyll-a, widely available from satellite ocean-colour observations, is commonly used as a measure of phytoplankton biomass but provides limited information on taxonomic composition. Accessory pigments, some of which are diagnostic of important phytoplankton groups, offer additional information on community structure, but their retrieval from ocean-colour data is challenging because of limited spectral resolution and strong covariance with chlorophyll-a. In this study, we evaluate machine learning methods for estimating diagnostic pigment concentrations from multispectral satellite observations. Using a global dataset of 33,640 High Performance Liquid Chromatography (HPLC) measurements matched with ESA Ocean Colour Climate Change Initiative (OC-CCI) reflectance data, we compare Random Forest and TabPFN models trained on multispectral reflectance with baseline models using chlorophyll-a alone. A temporally stratified validation scheme is employed to reduce the effects of autocorrelation. Results show that multispectral models consistently outperform approaches based solely on satellite-derived chlorophyll-a, demonstrating that ocean-colour reflectance contains additional information relevant to pigment discrimination. Improvements vary by pigment, with those strongly correlated with chlorophyll-a showing limited gains, while others exhibit substantial improvement. These findings highlight the potential of machine learning to extract ecologically relevant information from satellite data beyond conventional chlorophyll-based approaches.
David Moffat, Angus Laurenson, Victor Martinez-Vicente +4
Plymouth Marine Laboratory, Plymouth, UK · National Center for Earth Observation, Plymouth Marine Laboratory, UK · Department of Earth and Environmental Sciences, Centre for Geography and Environmental Science, University of Exeter, Cornwall, United Kingdom
Pseudo-nitzschia diatoms pose recurrent risks to coastal ecosystems and shellfish harvesting along the Portuguese Atlantic coast. Here we develop and evaluate a spatio-temporal machine-learning framework to predict harmful algal bloom (HAB) occurrence using exclusively satellite-derived predictors under realistic forecasting constraints. We characterised environmental and biological variability across shellfish production zones (L1-L9) using 5,882 observations, providing system-wide context. Predictive models were developed for zones L1-L2, a hotspot for Pseudo-nitzschia and domoic acid events, using a decade-long dataset (2013-2023; 1,440 observations; more than 1,000 satellite-based predictors including sea surface temperature, an upwelling index, chlorophyll-a, and plankton functional types). Sampling locations were partitioned into ecologically meaningful sub-regions using a river-aware spatial clustering scheme. A stringent spatio-temporal cross-validation strategy that simultaneously withholds entire years and spatial clusters prevents leakage and closely mimics real-world forecasting conditions. HAB occurrence proved moderately predictable across model classes and feature configurations. Ensemble tree-based methods achieved the strongest discrimination: Random Forest reached 0.74 +/- 0.05 with environmental predictors; Extra Trees reached 0.77 +/- 0.06 with biological variables added. Feature-importance analyses revealed that seasonal structure, spatial context, and lagged environmental conditions dominate model decisions, while biological indicators refine bloom likelihood within physically favourable periods. The framework demonstrates operationally relevant skill for satellite-supported HAB early-warning systems along eastern boundary upwelling coasts.
Ayman Bnoussaad, El Khalil Cherif, Ligia Pinto +3
ISR – Institute for Systems and Robotics, Lisbon, Portugal. · MARETEC/LARSyS, Associação do Instituto Superior Técnico para a Investigação e Desenvolvimento, Lisboa, Portugal. · IPMA – Portuguese Institute for Sea and Atmosphere, Algés, Portugal. +1
Data-driven models have advanced deterministic ocean forecasting, but extending machine learning to probabilistic global ocean prediction remains an open challenge. Here we introduce FuXi-ONS, the first machine-learning ensemble forecasting system for the global ocean, providing 5-day forecasts on a global 1° grid up to 365 days for sea-surface temperature, sea-surface height, subsurface temperature, salinity and ocean currents. Rather than relying on repeated integration of computationally expensive numerical models, FuXi-ONS learns physically structured perturbations and incorporates an atmospheric encoding module to stabilize long-range forecasts. Evaluated against GLORYS12 reanalysis, FuXi-ONS improves both ensemble-mean skill and probabilistic forecast quality relative to deterministic and noise-perturbed baselines, and shows competitive performance against established seasonal forecast references for SST and Niño3.4 variability, while running orders of magnitude faster than conventional ensemble systems. These results provide a strong example of machine learning advancing a core problem in ocean science, and establish a practical path toward efficient probabilistic ocean forecasting and climate risk assessment.
Qiusheng Huang, Xiaohui Zhong, Anboyu Guo +3
Artificial Intelligence Innovation and Incubation Institute, Fudan University, Shanghai, China. · Shanghai Innovation Institute, Shanghai, China. · Shanghai Academy of Artificial Intelligence for Science, Shanghai, China. +2