Source-linked AI summary
Beyond chlorophyll: machine learning estimates of diagnostic phytoplankton pigments from multispectral ocean colour data
David Moffat, Angus Laurenson, Victor Martinez-Vicente, Gemma Kulk, Xuerong Sun, Robert J. W. Brewin, Shubha Sathyendranath
TL;DR
Large-scale phytoplankton community composition is difficult to observe because chlorophyll-a provides limited taxonomic information and accessory-pigment retrieval is challenging. This study compares TabPFN and Random Forest models using multispectral ocean-colour data against a chlorophyll-a-only baseline. Multispectral models consistently outperform chlorophyll-only approaches, although gains vary among pigments and depend partly on chlorophyll-a covariance. The results support machine learning as a complementary way to derive ecologically meaningful pigment information from long-term ocean-colour records.
Problem
The study addresses whether multispectral ocean-colour observations contain information on phytoplankton pigment composition beyond chlorophyll-a, whose covariance with accessory pigments can confound apparent predictive skill.
Method
The study trains TabPFN and Random Forest models on multispectral ocean-colour observations matched to 33,640 HPLC measurements, using chlorophyll-a-only Random Forest models as a baseline.
Results
Multispectral models consistently outperformed chlorophyll-a-only approaches, with substantially larger improvements for alloxanthin, butfucoxanthin, and hex-butfucoxanthin.
Takeaways & Limitations
Machine learning can extract ecologically meaningful phytoplankton pigment-composition information from multispectral ocean-colour observations beyond chlorophyll-a.
Takeaways & Limitations
The relationships learned from past ocean-colour and pigment observations may change under future ocean conditions, requiring continued in situ measurements.
Abstract
from arXiv · showhide
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.
PREPRINT
This manuscript is an unreviewed ArXiv preprint submitted for peer review. Its content is preliminary and may change after review and publication.
- The manuscript is a non-peer-reviewed preprint submitted to ArXiv on 20 August 2026.
- The content has not undergone formal peer review and should therefore be considered preliminary.Subsequent versions may differ following peer review and publication.
1 INTRODUCTION
Phytoplankton are globally important, but their diverse community composition is difficult to quantify from large-scale observations. The study therefore tests whether multispectral ocean-colour data contain pigment-composition information beyond chlorophyll-a.
- Phytoplankton contribute around 50% of global primary production of organic carbon and form the marine food-web base.
- Identifying contributions from individual phytoplankton types is difficult because communities are highly diverse and existing in situ methods are expensive with limited geographic coverage.
- Accessory pigments provide information about phytoplankton groups, but pigment-to-group relationships are ambiguous and pigment proportions vary with physiology and environmental conditions.
- Prior work showed that predictor choice and cross-validation strategy can substantially affect estimated performance in phytoplankton-related machine-learning tasks.
- The study asks whether multispectral reflectance can discriminate accessory pigments beyond their covariance with chlorophyll-a.
2 MATERIALS & METHODS
The study combines HPLC pigment measurements with OC-CCI ocean-colour data, uses temporally held-out validation, and compares multispectral machine-learning models with a chlorophyll-a-only baseline.
- 2.1 In situ data: The dataset contains 33,640 HPLC measurements collected between 1991 and 2021, with quality filtering applied near the sensor detection limit.Values below 0.005 mg m−3 were removed when measurements were reported to three decimal places.
- 2.1 In situ data: Most pigments were strongly correlated with chlorophyll-a; fucoxanthin had r = 0.95, whereas zeaxanthin had r = 0.10.
- 2.2 Match-up data: HPLC observations were matched with same-day OC-CCI reflectances, Kd(490), and chlorophyll-a when samples fell within the satellite pixel and upper 5 m.
- 2.3 Data Partitioning: Four years—2002, 2007, 2012, and 2017—were withheld for validation to partially mitigate temporal autocorrelation and test generalisation to unseen years.
- 2.4 Machine learning: TabPFN and Random Forest models used multispectral reflectance, Kd(490), and chlorophyll-a, while the RF baseline used chlorophyll-a alone.
- 2.4 Machine learning: Models predicted eight pigments from log-transformed targets and were evaluated using R2, calculated in log-transformed space.
3 RESULTS
Multispectral models outperformed chlorophyll-a-only baselines on validation data, with performance and spectral gains varying across pigments. Global predictions showed coherent biogeographic patterns and differences among pigment distributions.
- Model performance: Validation R2 ranged from 0.23–0.74 for TabPFN, 0.13–0.72 for RF, and 0.075–0.48 for the chlorophyll-a-only baseline.Training R2 was higher across all model types, ranging from 0.89–0.97 for TabPFN, 0.87–0.95 for RF, and 0.83–0.95 for the baseline.
- Model performance: The chlorophyll-a-only baseline consistently underperformed TabPFN and RF models incorporating multispectral reflectance.TabPFN had the highest performance, while RF showed similar bias and root-mean-square difference.
- Pigment-specific differences: Fucoxanthin had high performance but the smallest improvement over the baseline, whereas 19’-butanoyloxyfucoxanthin and 19’-hexanoyloxyfucoxanthin showed the largest improvements.Fucoxanthin’s strong chlorophyll-a correlation helps explain its limited incremental gain from spectral inputs.
- Predictor interpretation: SHAP analysis showed chlorophyll-a dominated fucoxanthin and peridinin predictions, while spectral variables contributed substantially to other pigment models.The analysis compares the relative influence of chlorophyll-a, reflectance bands, and Kd(490) across RF models.
- Global distributions: Global pigment and pigment-to-chlorophyll-a maps reproduced coherent oceanographic structures and contrasting distributions among pigments.The maps include predicted concentrations and pigment-to-chlorophyll-a ratios generated with the spectral RF model.
4 DISCUSSION
The discussion concludes that multispectral reflectance provides information about pigment composition beyond chlorophyll-a, although gains differ among pigments. Explainable predictions and global patterns support ecologically meaningful retrievals, subject to data and temporal-scope limitations.
- Overall findings: Across all pigments, multispectral models outperformed chlorophyll-a-only models, with the magnitude of improvement varying substantially among pigments.The central comparison tests whether reflectance adds information beyond chlorophyll-a.
- Pigment-specific interpretation: Fucoxanthin had the strongest predictive performance but the smallest baseline improvement, consistent with its chlorophyll-a correlation of r = 0.95.SHAP analysis also showed chlorophyll-a dominated fucoxanthin predictions.
- Pigment-specific interpretation: Alloxanthin, 19’-butanoyloxyfucoxanthin, and 19’-hexanoyloxyfucoxanthin showed larger improvements, with reflectance bands and Kd(490) generally more influential than chlorophyll-a.These results support information about phytoplankton community composition beyond chlorophyll-a covariance.
- Pigment-specific interpretation: Zeaxanthin was least accurately predicted, yet its dependence on blue-green reflectance and distinct global distribution suggested an ecological signal beyond total phytoplankton biomass.The authors link this signal to cyanobacteria-associated pigments while noting that overall predictive accuracy remained modest.
- Ecological interpretation: Global maps showed biogeographic patterns consistent with contrasts between productive high-latitude waters and oligotrophic subtropical gyres.Fucoxanthin, peridinin, and alloxanthin were associated with larger phytoplankton in nutrient-rich regions, whereas zeaxanthin was associated with smaller phytoplankton.
- Implications: The approach offers a complementary way to derive phytoplankton community information from long-term ocean-colour records and supports future hyperspectral applications.The conclusion describes machine learning as complementary to conventional chlorophyll-based approaches.
- Limitations: Performance estimates may change under future ocean conditions because the algorithms rely on historical relationships between ocean colour and pigment measurements.Continued in situ HPLC measurements are identified as essential for improving algorithms and detecting climate-driven parameter changes.
5 CONCLUSIONS
Multispectral ocean-colour models estimate diagnostic pigments better than chlorophyll-only approaches, while revealing pigment-specific predictive skill and ecologically meaningful spatial patterns.
- Multispectral reflectance models consistently outperform chlorophyll-only approaches, indicating additional information about phytoplankton pigment composition.
- Predictive skill varies considerably among pigments, with fucoxanthin most accurate but strongly associated with chlorophyll-a.Its predictability may therefore derive substantially from covariance with phytoplankton biomass.
- Alloxanthin, butfucoxanthin and hex-butfucoxanthin show substantial improvement over chlorophyll-only models.
- SHAP signatures and global maps indicate that models capture pigment-specific variability rather than simply reproducing chlorophyll-a patterns.Different pigments correspond to distinct spectral and optical predictor combinations, with distributions consistent with known phytoplankton biogeography.
- Multispectral ocean-colour observations contain measurable diagnostic-pigment information beyond chlorophyll-a.Machine learning provides a framework for exploiting this information toward improved satellite monitoring of phytoplankton community composition.
FUNDING
The research received support from ESA through the BOOMS and BiCOME projects, alongside additional support from the Simons Foundation, ESA PHYTO-CCI, and UKRI.
- The research was supported by ESA through the BOOMS and BiCOME projects, with additional support from the Simons Foundation, ESA PHYTO-CCI and UKRI.