Machine Learning Approaches to Phytoplankton Identification and Classification Using GCOM-C/SGLI Imagery

Highlights

What are the main findings?

• Machine learning approaches were developed to classify phytoplankton types, including coccolithophores, diatoms, and dinoflagellates, using GCOM-C/SGLI satellite imagery.
• Random Forest (RF) and Gradient Tree Boosting (GTB) models outperformed other tested algorithms, achieving high accuracy of classification results.

What are the implications of the main findings?

• The developed machine learning models enable scalable monitoring of phytoplankton blooms, supporting both regional and global ocean observation.
• Combining remote sensing with artificial intelligence can be used as an alternative approach to monitor the marine ecosystem.

Abstract

Phytoplankton are fundamental to sustaining marine ecosystems and significantly influence the global carbon cycle. However, identifying their types accurately from satellite imagery remains a challenge. This study presents machine learning approaches for classifying phytoplankton types, including coccolithophores, diatoms, and dinoflagellates, using Second-generation Global Imager (SGLI) imagery aboard the GCOM-C satellite. Several algorithms were evaluated, with Random Forest (RF) and Gradient Tree Boosting (GTB) achieving the highest classification performance in classifying coccolitophores and diatoms. On the other hand, both RF and Classification and Regression Trees (CARTs) are effective for distinguishing dinoflagellates from surrounding water types. To assess model transferability, the developed machine learning models were applied in another sub-regions and on a different date of acquisition. The validation confirmed the ability of the model to generalize across sub-region and temporal variations in SGLI imagery. As a result, the potential of combined machine learning and SGLI imagery can improve phytoplankton detection, enabling large-scale monitoring at both regional and global levels. This paper highlights the importance of combining artificial intelligence with satellite-derived ocean color data to improve the monitoring of marine ecosystems.

1. Introduction

Phytoplankton, key primary producers in the ocean, generate about 50% of global net primary production and thereby shape nutrient cycling, carbon dynamics, and climate regulation [1,2,3]. Coccolithophores, diatoms, and dinoflagellates are important phytoplankton functional groups, each with distinct ecological and biogeochemical roles [4,5].

Coccolithophores are unicellular eukaryotic phytoplankton that contribute significantly to the global carbon cycle through particulate inorganic carbon (PIC) production and calcification. Coccolithophores produce calcium carbonate (CaCO₃) plates (coccoliths) that regulate seawater optics and sustain the oceanic carbonate pump via calcification flux [6,7,8]. During blooms, detached coccoliths in suspension intensify backscattering and reflectance at times forming striking turquoise patches visible in satellite images [9,10]. These features complicate discrimination from non-calcifying taxa and suspended particulates [4,11]. Diatoms, by contrast, build siliceous frustules and dominate in high-nutrient, upwelling, or high-latitude regimes; they are among the most efficient drivers of organic carbon export (the biological pump) in many ocean regions [12,13,14]. Dinoflagellates are prominent drivers of harmful algal blooms (HABs) that threaten marine resources and ecosystem stability, with trophic flexibility spanning autotrophy, heterotrophy, and mixotrophy [15,16,17]. Due to this diversity in physiological and optical traits, distinguishing these groups remotely is crucial for understanding ecosystem dynamics and carbon cycling.

Satellite ocean color remote sensing has been widely used to monitor phytoplankton distributions since the 1990s, via sensors such as SeaWiFS, MODIS, MERIS, and OLCI [18,19,20]. These platforms measure water-leaving reflectance, which is related to inherent optical properties (absorption and scattering) of water constituents. However, reliably discriminating phytoplankton functional types (PFTs) via conventional empirical or semi-analytical algorithms remains elusive. Coccolithophore optical signals often overlap with those from non-alkalinity variance and suspended sediments; traditional band-ratio or threshold-based indices yield ambiguous results in optically complex waters, leading to misclassification [4,8]. On the other hand, diatoms and dinoflagellates share overlapping pigment absorption features and scattering characteristics with other taxa, further confounding discrimination [21]. In coastal or turbid waters, high CDOM or suspended particulate load further degrades classification performance [22,23,24].

To address these limitations, machine learning (ML) has emerged as a promising path forward. Machine learning (ML) methods can be used to capture complex nonlinear relationships between spectral signatures and phytoplankton groups, which offers high performance on high-dimensional and noisy datasets [25,26]. Recent reviews highlight the growing integration of ML in ocean color retrieval, classification, and fusion tasks [27]. Machine Learning can be used to rapidly retrieve ocean color information from hyperspectral satellite observations under the influence of clouds, aerosols, and sunglint [28]. In addition, machine learning has proven to outperform traditional physics- and statistics-based algorithms in image information extraction across various domains, and its potential is increasingly being recognized in ocean remote sensing applications [29]. On the other hand, the most effective artificial neural network architectures for estimating chlorophyll concentrations in coastal waters using data from the MODIS Aqua ocean color sensor have been identified [30]. Several algorithms have been developed to estimate the large-scale distribution of phytoplankton cell sizes, referred to as phytoplankton size classes (PSCs), in surface ocean waters using remotely sensed data. Among the tested methods, the Random Forest approach outperformed PLS, ANN, and SVM in calibrating PSC retrieval models [31].

The GCOM-C (Shikisai) satellite mission, operated by JAXA, carries the Second-generation Global Imager (SGLI) sensor, offering 19 spectral bands across visible to shortwave infrared wavelengths and spatial resolutions from 250 m to 1000 m [32,33,34]. To maximize the potential of SGLI, a shift from conventional thresholds to a data-driven classification framework is required.

This study aims to develop a machine learning model to identify and classify coccolithophores, diatoms, and dinoflagellates using GCOM-C/SGLI imagery. Several machine learning algorithms will be evaluated and compared with particular emphasis on ensemble classifiers that have demonstrated high predictive performance. This study will also assess the developed machine learning models for another sub-region and on a different date of acquisition. The aim is to demonstrate the ability of the models to generalize across sub-regions and temporal variations in SGLI imagery. The implications of these findings are for operational monitoring of marine ecosystems.

2. Materials and Methods

2.1. Site Selection and Rationale

To demonstrate the reliability of the proposed method and ensure broader applicability, SGLI data from several sub-regions were utilised in this study, as illustrated in Figure 1. We also used NOAA (National Oceanic and Atmospheric Administration) ETOPO1 which provides a 1-arc-minute resolution representation of Earth’s relief, integrating both land elevations and ocean-floor bathymetry. Several existing studies have identified or confirmed the presence of optical water and phytoplankton types in different sub-regions, as described in

Table 1. Descriptions of the expected or confirmed optical water and phytoplankton types in different sub-regions.

Sub-Region	Observation Date	Consideration/Confirmation
Sagami Bay	17 May 2020	Confirmed phytoplankton coccolithophore bloom [35]
Waters of Southeast Hokkaido	13 October 2021	Confirmed red tide caused by phytoplankton from the dinoflagellate group [36,37]
	8 May 2022	Expected to be seasonal blooms of phytoplankton diatoms
	7 May 2022	Expected to be seasonal blooms of phytoplankton diatoms

, which were used to verify the presence of phytoplankton in these areas. This is to assist in creating datasets for a machine learning model.

2.2. GCOM-C/SGLI Data Acquisition and Extraction

The GCOM-C/SGLI provides 19 spectral bands at spatial resolutions ranging from 250 m to 1000 m, with a global revisit time of approximately two days. Its unique bands, such as 380 nm UV and shortwave infrared, enhance its capacity to resolve atmospheric and oceanic processes. Level-2 SGLI products underwent atmospheric correction using JAXA’s standard algorithms, followed by cloud masking, radiometric calibration, and sunglint removal. Atmospheric correction was conducted using standard SGLI algorithms, ensuring removal of aerosols and Rayleigh scattering. Cloud and glint pixels were masked. The reflectance spectra were normalised to minimise illumination differences.

This study used the SGLI Level-2 dataset with the latest Version 3 atmospheric correction that includes remote sensing reflectance (Rrs, sr⁻¹) and Chl-a at a spatial resolution of 250 m [38]. The Second-generation Global Imager (SGLI) measures remote-sensing reflectance (Rrs) at seven discrete wavelengths (380, 412, 443, 490, 530, 565, and 670 nm) spanning the ultraviolet to visible spectral range. Chl is usually used to identify the presence of phytoplankton, but it cannot be used to accurately determine the type of phytoplankton. Therefore, both Rrs and Chl were used to identify the type of phytoplankton for machine learning datasets.

2.3. Machine Learning Approach to Phytoplankton Identification and Classification

Phytoplankton classification using GCOM-C/SGLI imagery and machine learning is presented in Figure 1. All image preprocessing, classification and analysis were conducted using the Google Earth Engine platform [39]. A crucial aspect of implementing machine learning is the construction of an appropriate dataset, as it enables the model to learn and effectively distinguish the target classes. In this study, we employed remote sensing reflectance (R_rs) at discrete wavelengths (380, 412, 443, 490, 530, 565, and 670 nm), hereafter denoted as b1–b7. Chlorophyll concentration values were incorporated to guide class definition, while published phytoplankton data were additionally utilized to ensure the dataset accurately represented each category. The dataset was divided into two parts: a training dataset for the machine learning process (70%) and a testing dataset to evaluate the reliability of the machine learning model (30%). We also used bathymetry data (NOAA ETOPO1) to enhance classification accuracy, particularly in distinguishing between classes such as turbid waters and coccolithophore-dominated waters. The final model can be used to identify and classify phytoplankton in other areas or other acquisition dates of the GCOM-C/SGLI imagery.

This study used several machine learning methods to identify and classify phytoplankton. We used Random Forest (RF), Classification, Regression Tree (CART), and Gradient Tree Boosting (GTB). Tree-based algorithms especially RF have consistently demonstrated superior performance in remote sensing applications compared to traditional statistical and classical machine learning classifiers, such as Maximum Likelihood and Support Vector Machines (SVMs) [40,41,42,43]. They have consistently demonstrated superior performance in remote sensing classification tasks. Unlike parametric classifiers, these methods do not rely on distributional assumptions of spectral features and are therefore more robust when applied to heterogeneous environments. RF is widely recognized for its robustness and computational efficiency, while GTB has shown even higher accuracies in distinguishing spectrally overlapping classes, such as turbid waters and coccolithophore blooms. Compared to a classical machine learning algorithm (e.g., ANN, SVM, or kNN), these algorithms not only provide improved accuracy but also enhanced generalization capacity, making them particularly suitable for complex ocean-color and phytoplankton classification problems.

The RF algorithm was employed in this study as the primary machine learning method due to its robustness and suitability for both classification and regression tasks [44]. RF is an ensemble learning technique that combines the predictions of multiple decision trees to improve generalization performance and reduce overfitting. This method is based on the principle of bootstrap aggregating, whereby multiple subsets of the training dataset are generated through random sampling with replacement. Each subset is used to train an individual decision tree, thereby introducing diversity among the base learners.

During the construction of each tree, a random subset of features is selected at every decision node rather than considering the entire feature set. This randomization ensures a low correlation among trees, further enhancing the ensemble’s predictive power. Each decision tree is grown to its maximum extent without pruning, which allows the trees to capture complex patterns in the data. For classification tasks, the final prediction is determined through majority voting across all trees. Mathematically, the Random Forest prediction for classification can be expressed as:

$$y=mode\{h_1\left(x\right),h_2\left(x\right),h_3\left(x\right),\dots ,h_B\left(x\right),\}$$

(1)

where $h_B\left(x\right)$ denotes the prediction of the $b^{th}$ decision tree and $B$ represents the total number of trees in the forest. Several hyperparameters govern the performance of the Random Forest model. The number of trees ($\mathit{n\_estimators}$) influences the stability and accuracy of the predictions, while the maximum number of features considered at each split ($\mathit{max\_features}$) controls the diversity among trees. Tree complexity is managed through parameters such as maximum depth ($\mathit{max\_depth}$) and the minimum number of samples per split or leaf, which serve as stopping criteria during tree construction.

The RF algorithm offers several advantages, including resilience to noisy data and outliers, the capacity to handle high-dimensional feature spaces, and the provision of internal estimates of feature importance, which enhance interpretability.

CART represents a non-parametric supervised machine learning approach that partitions input data into homogeneous subsets [45]. The final model is structured as a binary tree in which terminal nodes correspond to predicted classes or continuous values. CART offers simplicity, interpretability, and direct feature selection, but its predictive accuracy is often constrained by high variance and sensitivity to noise, particularly in heterogeneous remote sensing environments.

On the other hand, GTB is an advanced ensemble learning approach that constructs predictive models by sequentially combining multiple decision trees, where each tree is trained to correct the residual errors of the preceding ones [46]. GTB has become a widely adopted technique due to its ability to handle high-dimensional, heterogeneous, and often noisy spectral and spatial features. Therefore, this classifier is suitable to address complex ocean-color and phytoplankton classification issues.

2.4. Assessment of the Results

To ensure the robustness and generalizability of the model, the performance of the classification model was assessed using a confusion matrix, which provides a comprehensive comparison between predicted and actual class labels. The matrix summarizes classification outcomes in terms of correctly and incorrectly classified samples, structured as true positives, true negatives, false positives, and false negatives. From this matrix, several evaluation metrics were derived, namely Producer’s Accuracy, User’s Accuracy, Overall Accuracy, Kappa Coefficient, Correlation Coefficient and R² which together provide a rigorous assessment of classification reliability [47].

Producer’s Accuracy ($PA$) represents the probability that a reference sample of a given class is correctly classified. It is calculated as the ratio of correctly classified samples in a particular class to the total number of reference samples for that class. $PA$ is analogous to recall or sensitivity, and it reflects how well the classifier can recognize members of a given class:

$$P{A}_i={\displaystyle \frac{n_{ii}}{{\sum }_{j=1}^k{n}_{ij}}}$$

(2)

User’s Accuracy ($UA$) measures the probability that a sample classified into a given class actually belongs to that class. It is calculated as the ratio of correctly classified samples in a class to the total number of samples assigned to that class. $UA$ is equivalent to precision, and it indicates the reliability of the classification from the user’s perspective:

$$U{A}_i={\displaystyle \frac{n_{ii}}{{\sum }_{j=1}^k{n}_{ji}}}$$

(3)

Overall Accuracy ($OA$) provides a general measure of classification performance by dividing the total number of correctly classified samples by the total number of samples in the dataset:

$$OA={\displaystyle \frac{{\sum }_{j=1}^k{n}_{ii}}{N}}$$

(4)

where $N$ is the total number of samples across all classes. While $OA$ is widely used, it may not fully capture classification reliability in the presence of class imbalance. To address this limitation, the $Kappa Coefficient \left(\kappa \right)$ was calculated. Kappa measures the agreement between the classification results and the reference data, adjusted for the agreement that could occur by chance. It is expressed as:

$$\kappa ={\displaystyle \frac{OA-P_e}{1-P_e}}$$

(5)

where $P_e$ is the hypothetical probability of chance agreement, calculated from the row and column totals of the confusion matrix.

In this study, we used Pearson correlation that compares actual class and predicted class. It shows whether the classifier’s predictions increase or decrease in a pattern similar to the actual labels. To classify the correlation coefficients as indicating weak, moderate, or strong associations, we applied the interpretive thresholds outlined in

Table 2. Interpretation of Correlation Coefficient [48].

Absolute Magnitude of Correlation Coefficient	Interpretation
0.00–0.10	Negligible correlation
0.10–0.39	Weak correlation
0.40–0.69	Moderate correlation
0.70–0.89	Strong correlation
0.90–1.00	Very strong correlation

[48].

3. Results

The identification and classification of phytoplankton using GCOM-C/SGLI imagery are divided by phytoplankton types, i.e., coccolithophores, diatoms, and dinoflagellates. Each type of phytoplankton in this study will be identified using several machine learning methods in this section and analysed to find the best machine learning model. In our study, we adopted a simpler 70/30 train–test split approach to maintain computational efficiency and ensure a clear separation between the training and independent testing datasets. Given the large size and representativeness of our dataset, this split was sufficient to achieve stable and reliable model performance without significant overfitting.

We conducted various scenarios with different inputs for these ML models and tested them to determine the optimal scenario. The scenarios are utilising: (1) seven bands of SGLI data using RF, (2) six bands of SGLI data excluding Rrs₃₈₀ using RF, and (3) seven bands of SGLI data using CART, and (4) seven bands of SGLI data using GTB.

3.1. Coccolithophore Blooms Identification and Classification

Three classifiers, i.e., RF, CART, and GBT, were tested to identify and classify coccolithophore blooms in Sagami Bay, Japan. The following are the results of coccolithophore blooms identification and classification using various scenarios with three different ML classifiers. We used a 70/30 train–test split approach with 203 samples for the coccolithophore class, 57 samples for the turbid class, 86 samples for the oligotrophic class, 26 samples for the mesotrophic class, and 29 samples for the mix water class. The spatial distribution of all samples were spread evenly across each class and split for 70/30 train–test randomly.

Figure 2a,b show the Rrs values of each class provided from the datasets and the average Rrs values for each class, with Rrs plotted against different spectral bands (B1 to B7), respectively. The turbid water demonstrates the strongest reflectance in almost all bands. It reaches a peak around B5 and B6, which indicates a high concentration of suspended particles that scatter light. The presence of coccolithophores can also be identified by their high reflectance in the visible band.

We can visually inspect the classification results in Figure 3, where all classifiers accurately classify each class. All classifiers obtained high accuracy results, with their accuracy varying across scenarios (

Table 3. Accuracies of the coccolithophore classification results.

Scenario	Overall Accuracy	Kappa Coefficient	Pearson Correlation Coefficient	Determination Coefficient (R2)
RF with seven bands	0.976	0.962	0.977	0.954
RF with six bands	0.939	0.911	0.931	0.861
CART with seven bands	0.967	0.950	0.956	0.908
GTB with seven bands	0.984	0.975	0.987	0.974

). In the seven-bands scenario, the GBT classifier achieved the best result, with an overall accuracy of 0.984 (Kappa = 0.975), followed by the RF, which achieved an accuracy of 0.976 (Kappa = 0.962). CART also produced competitive results (0.967, Kappa = 0.950), although its classification threshold was less precise.

The results showed that excluding band 1 (Rrs₃₈₀) made a significant decrease in accuracy, as tested in RF (Accuracy = 0.939, Kappa = 0.911), highlighting the importance of the Rrs₃₈₀ for detecting coccolithophores. This finding is consistent with previous research highlighting how coccolithophore blooms, which are rich in calcite plates, exhibit strong backscatter signals at shorter wavelengths.

To validate the proposed method, we tested the ML model obtained in the previous process in another sub-region in the south of Plymouth City in the English Channel area. As shown visually in Figure 4, the classification result demonstrates high accuracy. However, there is an oddity in the detection of coccolithophores in deep-sea regions due to the detected turbidity, as turbidity is typically caused by suspended particles, such as mud, clay, and organic matter, in the water and is more commonly found in shallow waters. This anomaly was verified using chlorophyll data from GCOM-C/SGLI, which showed that the chlorophyll concentration in the area was extremely high. This indicates that the area is a coccolithophore because the phytoplankton were also detected in the surrounding area. To address this issue, the proposed method utilised bathymetry data from ETOPO1 to separate the turbid and coccolithophore classes, thereby achieving more accurate classification results. It shows promising results because the deep-sea area that was initially classified as turbid has been reclassified as the correct class, which is what it should be.

3.2. Diatom Blooms Identification and Classification

The following are the results of diatom blooms identification and classification using various scenarios with three different ML classifiers. Figure 5 presents the average remote sensing reflectance (Rrs) spectra across different air classes—Diatom, Low Diatom, Turbid, and Clear—plotted against b1 to b7, the spectral bands. Across all classes, Rrs values are lowest at b1 (shorter wavelengths), gradually increasing towards the b2 to b4 visible spectrum, before decreasing again at longer wavelengths (b6 to b7). Diatom and Low-Diatom waters show enhanced reflectance in b3 to b6 due to phytoplankton pigment absorption, dominated by chlorophyll-a near b4 to b5. Turbid waters show the highest reflectance across all bands, which reflects the strong backscattering caused by suspended particles that is characteristic of turbid conditions.

Diatom classification using GCOM-C SGLI images and machine learning algorithms showed high performance across all tested scenarios (

Table 4. Accuracies of the diatom classification results.

Scenario	Overall Accuracy	Kappa Coefficient	Pearson Correlation Coefficient	Determination Coefficient (R2)
RF with seven bands	0.978	0.969	0.970	0.940
RF with six bands	0.988	0.983	0.996	0.991
CART with seven bands	0.966	0.953	0.966	0.933
GTB with seven bands	0.978	0.969	0.970	0.940

, Figure 6). The RF and GTB model with seven spectral bands achieved the highest accuracy at 0.978 (Kappa = 0.969). The RF model trained with six bands produced slightly lower but still robust results (Overall Accuracy = 0.964, Kappa = 0.950). The CART algorithm using seven bands performed comparably, with an overall accuracy of 0.966 (Kappa = 0.953).

These findings indicate that including all seven spectral bands yields high accuracy; on the other hand, removing b1 results in a measurable increase in performance. Among the classifiers, the RF and GTB have outperformed the single tree approach (CART) in handling the spectral variability of diatoms. Classification results show that diatom identification using GCOM-C SGLI is feasible with high reliability, particularly when machine learning methods are employed and complete spectral information is used.

We validated the developed RF model with seven bands of SGLI data from a different acquisition date to assess its reliability. The classification result in Figure 7 shows that it was successful in identifying diatoms, aligning with the original training dataset. While minor discrepancies were found, the validation confirmed the ability of the model to generalize across temporal variations in SGLI imagery.

3.3. Dinoflagellate Blooms Identification and Classification

The following are the results of dinoflagellate blooms identification and classification using various scenarios with three different ML classifiers. Figure 8 shows the average Rrs spectra across seven spectral bands for six distinct water classes, including harmful algal bloom (HAB) cases (Karenia and Karenia-low), mixed phytoplankton assemblages (Mix-low and Mix-high), turbid waters, and clear waters. The Karenia and Karenia-low classes have low reflectance across all bands, particularly in the b1 to b4 visible range. This spectral depression reflects the high chlorophyll-a and accessory pigment concentrations typical of dense Karenia blooms, which dominate light absorption and suppress water-leaving radiance.

The RF, CART, and GTB classifiers with seven bands generated bloom detections that closely matched the SGLI baseline, with minimal misclassification and clear delineation of bloom boundaries. Dinoflagellate blooms were identified and classified using GCOM-C SGLI data combined with RF, CHART, and GTB (Figure 9). Overall accuracies and Kappa coefficients for the tested scenarios are presented in

Table 5. Accuracies of the dinoflagellate classification results.

Scenario	Overall Accuracy	Kappa Coefficient	Pearson Correlation Coefficient	Determination Coefficient (R2)
RF with seven bands	0.988	0.982	0.997	0.994
RF with six bands	0.922	0.888	0.928	0.846
CART with seven bands	0.988	0.982	0.997	0.994
GTB with seven bands	0.976	0.963	0.994	0.987

. The RF and CART classifiers using seven spectral bands produced the highest accuracies, with an overall accuracy of 0.988 and corresponding Kappa coefficients of 0.982. In contrast, reducing the RF input to six bands resulted in a decline in performance, with an overall accuracy of 0.922 and a Kappa coefficient of 0.888. GTB with seven bands also demonstrated strong performance but was slightly less accurate, yielding an overall accuracy of 0.976 and a Kappa coefficient of 0.963.

4. Discussion

Our findings reinforce the growing evidence that machine learning (ML) methods, particularly ensemble approaches, are effective for classifying coccolithophore, diatom, and dinoflagellate blooms from SGLI images. Consistent with previous studies, RF and GTB achieved superior performance relative to a classical machine learning algorithm, such as SVM [49,50]. This finding also aligns with a prior study that RF outperformed other ML approaches such as ANN, MLR, SVR and XGBoost [31,51]. The comparatively weak performance of CART is likely attributable to its limited capacity to generalize from a relatively small training dataset, as it is a single decision tree model.

Model transferability tests highlight the importance of incorporating environmental context alongside spatial and spectral information. When the RF coccolithophore model was applied to a different sub-region in the English Channel, the inclusion of bathymetric information improved classification accuracy, suggesting that coupling SGLI reflectance with geophysical parameters enhances robustness in heterogeneous coastal and open-ocean environments. This finding aligns with prior work emphasizing the spatial and spectral transferability of machine learning models [52,53,54].

Temporal validation further demonstrated that the RF model performed consistently across different acquisition dates, indicating relative stability in diatom spectral signatures despite seasonal or environmental variability. Dinoflagellate results demonstrate that SGLI’s spectral resolution is sufficient for operational HAB monitoring. Both RF and CART achieved strong classification accuracy, reinforcing earlier findings that SGLI can reliably distinguish dinoflagellates from co-occurring water types [36,55,56]. This capability supports applications in early-warning systems and routine ecological assessments.

Overall, these results position our work within a growing body of literature demonstrating the advantages of ensemble ML approaches using SGLI images for phytoplankton classification, reinforcing the prior works that ML can identify phytoplankton using ocean colour data [57,58]. By evaluating three phytoplankton groups and explicitly testing spatial and temporal transferability, our study extends the current literature on operational marine-ecosystem monitoring and provides evidence for the broader applicability of ML-based classification frameworks.

5. Conclusions

This study demonstrated the effectiveness of machine learning models, particularly Random Forest and Gradient Tree Boosting, in classifying phytoplankton, including coccolithophores, diatoms, and dinoflagellates, using GCOM-C/SGLI imagery. These findings could inform an alternative approach to operational monitoring of phytoplankton on regional and global scales.

Despite encouraging results, some limitations warrant consideration. First, classification accuracies may vary under conditions of optically complex waters, where non-algal particles and dissolved organic matter confound spectral signals. Validation with in situ datasets across diverse aquatic environments is needed to confirm model generalizability. Second, while GCOM-C SGLI provides sufficient spectral coverage for diatom detection, hyperspectral sensors such as PACE OCI could further improve discrimination of phytoplankton functional types. Future work should also explore temporal dynamics, including seasonal blooms, to assess the potential of SGLI for long-term ecosystem monitoring. A comparative analysis with existing PFT algorithms is essential to identify the most effective approach, which will be addressed in future work.