Development of Parameter-Tuned Algorithms for Chlorophyll-a Concentration Estimates in the Southern Ocean

Highlights

What are the main findings?

• Satellite-derived Chlorophyll-a (Chl) concentration products (MODIS, OC-CCI, and GlobColour) systematically underestimated Chl concentrations in the Southern Ocean (SO), particularly at high concentrations (Chl > 0.2/0.3 mg/m³).
• Developed parameter-tuned OC3M-based algorithms substantially improved Chl retrieval accuracy, with R² increasing to 0.68–0.91, slopes approaching 1.0 (0.62–0.92), and notable reductions in MAE (1.39–1.42) and RMSE (1.49–1.51).

What are the implications of the main findings?

• A high-precision (high-performance liquid chromatography-derived, HPLC), long-term (1997–2021), and spatially extensive (south of 40°S) in situ Chl dataset was used for the first time to enhance satellite-based Chl retrieval accuracy in the data-sparse and optically complex SO.
• These findings provide a foundation for developing regional algorithms and advancing long-term, large-scale assessments of phytoplankton dynamics.

Abstract

Accurate estimates of Chlorophyll-a (Chl) concentration from satellite observations are critical for understanding large-scale phytoplankton variations, particularly in the context of climate change. However, existing operational Chl retrieval algorithms have been shown to perform poorly in the Southern Ocean (SO). To address this issue, this study proposed improved Chl algorithms tailored to the SO. To this end, three Chl satellite products (MODIS, OC-CCI, and GlobColour) were evaluated against high-precision (high-performance liquid chromatography-derived, HPLC), long-term (1997–2021), and spatially widespread (south of 40°S) in situ Chl observations. Subsequently, OC3M-based empirical algorithms were improved using remote sensing reflectance (R_rs) data. Among the original products, OC-CCI exhibited the best overall performance (R² = 0.36, Slope = 0.36), followed by GlobColour-AVW (R² = 0.27, Slope = 0.21), whereas Aqua-MODIS showed the worst agreement (R² = 0.18, Slope = 0.18) with in situ observations. All three products systematically underestimated Chl concentrations, with average biases of 43% (Aqua-MODIS), 24% (OC-CCI), and 36% (GlobColour-AVW), particularly at high Chl concentrations (>0.2 mg/m³ for Aqua-MODIS and GlobColour-AVW; >0.3 mg/m³ for OC-CCI). The parameter-tuned algorithms significantly reduced these biases to 1% (OC-CCI), 3% (GlobColour-AVW), and a slight overestimation of 2% (Aqua-MODIS). All products showed marked improvements in performance, with R² increasing to 0.68–0.91, slopes approaching 1.0 (0.62–0.92), and notable reductions in MAE (1.39–1.42) and RMSE (1.49–1.51). These results offer enhanced capabilities for Chl retrieval in the data-sparse and optically complex waters of the SO.

1. Introduction

The Southern Ocean (SO), the oceanic region between 30°S and Antarctica, despite covering only ~30% of the global ocean surface, plays a vital role in the Earth’s climate system [1]. It acts as both a major carbon sink and a key heat regulator, transporting heat through ocean currents and regulating global temperature distribution [2,3,4,5]. Over 40% of the global ocean’s carbon uptake and 50% of its heat uptake occurred in the SO between the decades of 1851–1860 and 2005–2014, significantly mitigating climate change [6]. However, ongoing atmospheric and oceanic warming has rendered the SO highly vulnerable to climate change, threatening its important role in regulating global energy balance and biogeochemical cycles [7,8,9]. Therefore, large spatiotemporal coverage and continuous monitoring of the SO are crucial for understanding and responding to climate change [10].

In the surface waters of the SO, phytoplankton photosynthesis transforms carbon dioxide into new particles and dissolved organic carbon, which is crucial to the global carbon cycle [11]. Chlorophyll-a (Chl), the key pigment for phytoplankton photosynthesis, serves as an essential indicator for assessing marine ecosystem health and productivity [12,13]. Since the pioneering measurements of Chl concentration [14], Chl has been widely used as a proxy for phytoplankton biomass. However, in situ Chl collections in the SO are often constrained due to harsh polar conditions and remote distances. Fortunately, satellite ocean colour remote sensing, with its large-scale and long-term observation capabilities, provides an effective tool for understanding ocean ecology and biogeochemistry on synoptic scales [15,16,17]. Despite this, satellite observations in polar regions are limited by low-to-non-existent light during winter months, as well as high sea ice and cloud cover [18]. These factors are particularly challenging in the SO, where unique bio-optical properties further affect the performance of Chl inversion algorithms [19,20,21,22,23].

Current satellite Chl retrieval algorithms have been reported to perform poorly in the SO, underestimating Chl concentrations by up to ~70% [19,21,24,25]. This underperformance is not yet fully understood, but likely contributing factors include pigment composition and packaging [21], particle backscattering [26], low coloured dissolved organic matter [23] and sea ice contamination [20]. These factors can modify the spectral signatures captured by ocean colour sensors, thereby introducing biases in Chl retrieval. To improve the accuracy of satellite Chl retrievals, several regional algorithms have been developed for the SO, such as OC4Sze [27], OC4Jo&GLOJo [28], OC3M/FURG-SO [29], and OC4-SO [24].

Several limitations constrain the applicability of these algorithms in the SO. First, many studies relied partly or entirely on in situ Chl measurements obtained from fluorometry rather than the high-performance liquid chromatography (HPLC) method [27], despite evidence that HPLC provides higher accuracy [30,31,32]. In addition, most algorithms were originally developed using data from a specific sensor, such as the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) and the Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua. However, long-term ocean colour datasets (e.g., ESA OC-CCI and CMEMS GlobColour) are typically constructed from merged observations collected by multiple sensors. Thus, regional algorithms should incorporate multi-sensor merged data, as demonstrated by Ferreira et al. [24], who developed the OC4-SO algorithm for the western Antarctic Peninsula using the OC-CCI product. Moreover, existing algorithms are based on limited and outdated in situ Chl datasets, mainly collected near the Antarctic Peninsula. For example, Johnson et al. [28] developed an OC4v6-based algorithm using data from 2001 to 2008, and Pereira et al. [29] proposed an OC3M-based algorithm for the northern Antarctic Peninsula using data from 2002 to 2010. Consequently, there remains a lack of robust algorithms for the SO that integrate high-accuracy (HPLC-derived) and long-term (>20 years) in situ observations with multi-source merged satellite products.

To address the limitations of existing models, this study aims to develop parameter-tuned Chl retrieval algorithms for the SO. The specific objectives are: (i) quantifying the accuracy of current satellite Chl products (i.e., Aqua-MODIS, OC-CCI and GlobColour) using the most extensive and spatially comprehensive in situ Chl dataset available for the SO; (ii) developing new parameter-tuned Chl algorithms based on remote sensing reflectance (R_rs), including multi-sensor merged datasets; and (iii) assessing the performance of the proposed algorithms using independent in situ Chl measurements. The findings will provide valuable support for Chl inversion in the complex polar ocean.

2. Materials and Methods

2.1. In Situ Data

A total of 3836 Chl measurements were obtained from field sampling carried out in the SO, all derived through HPLC. This dataset included 3556 quality-controlled records published by Valente et al. [33] and an additional 280 surface samples collected during two research voyages: the Antarctic Circumnavigation Expedition (ACE) (https://swisspolar.ch/expeditions/ace/; accessed on 10 May 2024) [34] and the Southern Ocean Large Area Carbon Export (SOLACE) (https://aappartnership.org.au/solace/; accessed on 10 May 2024) expedition. In situ Chl was quantified as the total concentration of mono- and divinyl chlorophyll-a and chlorophyllide-a, as well as the allomeric and epimeric forms of Chl, when available [22].

The dataset compiled by Valente et al. [33], available through the OC-CCI project v3.0 (https://doi.org/10.1594/PANGAEA.941318; accessed on 10 May 2024), includes global bio-optical in situ observations. It merged bio-optical data from 27 data sources and covered the period from 1997 to 2021. To ensure the higher quality and reliability of the measured data, only HPLC-based Chl measurements collected south of 40°S were used in this study. From the ACE, 221 surface observations were collected aboard the RV Akademik Tryoshnikov during the Austral Summer from 20 December 2016 to 19 March 2017. The SOLACE contributed 59 surface observations, collected aboard the Australia’s RV Investigator during the Austral Summer from 7 December 2020 to 31 January 2021.

Combined, a dataset of 3836 in situ Chl measurements spanning from 13 October 1997 to 13 January 2021, covering 40°S to 78°S, were collected in this study (

Table 1. Summary of in situ dataset.

Data Source *	Date	Chl Range (mg/m3)	Number of Observations
AMT	15 October 1997–27 October 2018	0.033–4.460	66
AWI	20 April 2008–13 May 2018	0.083–5.554	504
CC	21 May 2007–29 May 2007	0.171–1.527	21
IMOS	15 September 2014–22 March 2017	0.095–1.303	49
MAREDAT	1 November 1997–20 February 2006	0.005–9.870	229
NOMAD	14 October 1997–6 March 2007	0.031–6.303	130
PALMER	15 December 1997–8 February 2016	0.010–22.689	1350
SeaBASS	13 October 1997–12 January 2017	0.005–7.532	1201
TARA	29 November 2010–27 January 2011	0.099–1.422	6
ACE	23 December 2016–16 March 2017	0.016–3.733	221
SOLACE	7 December 2020–13 January 2021	0.092–1.099	59

* AMT: Atlantic Meridional Transect; AWI: data collection from Astrid Bracher; CC: COASTCOLOUR; IMOS: Australia’s Integrated Marine Observing System; MAREDAT: MARineEcosytem DATa; NOMAD: NASA bio-Optical Marine Algorithm Data set; PALMER: Palmer Station Long-term Ecological Research; SeaBASS: SeaWiFS Bio-optical Archive and Storage System; TARA: data collection from global transects; ACE: Antarctic Circumnavigation Expedition; SOLACE: Southern Ocean Large Area Carbon Export.

and Figure 1). Chl concentrations ranged from 0.005 mg/m³ to 22.689 mg/m³, with an average of 1.174 mg/m³.

2.2. Satellite Data

Satellite-derived Chl concentration and R_rs products were extracted from three sources: the NASA MODIS on the Aqua satellite (hereafter Aqua-MODIS, [35]), the ESA Ocean-Color Climate Change Initiative (OC-CCI, [36]), and the Copernicus Marine Environmental Service (CMEMS) GlobColour project [37]. All datasets used in this study were Level-3 binned (L3b) 8-day composite products with a spatial resolution of 4 km. The Integerised SINusoidal (ISIN) projection was adopted, as it has been shown to perform better in high-latitude regions [38]. To balance temporal resolution and data availability, 8-day composite products were selected, although they may somewhat obscure short-term variability [28,39,40]. Daily products, however, often suffer from limited coverage, leading to an insufficient number of valid match-ups for robust statistical analysis. To further assess the potential influence of this choice, a sensitivity experiment comparing daily and 8-day products is provided in Appendix A (Figure A1 and

Table A1. Statistical metrics for OC-CCI product versus in situ data.

	Match-Ups	Bias	MAE	RMSE	R2	Slope	Intercept
Daily	408	0.81	1.69	1.97	0.65	0.47	0.25
8-Day	1329	0.76	1.84	2.27	0.36	0.36	0.31
Difference	921	0.05	0.15	0.30	0.29	0.11	0.06

).

To align with the collection period of in situ Chl (Table 1), Aqua-MODIS data (https://oceancolor.gsfc.nasa.gov/; accessed on 10 May 2024) spanned from 28 July 2002 to 16 January 2021, OC-CCI v6.0 datasets (http://www.esa-oceancolor-cci.org; accessed on 10 May 2024) covered the period from 29 August 1997 to 17 January 2021, and GlobColour version 4.2.1 (https://hermes.acri.fr/; accessed on 10 May 2024) covered the period from 29 August 1997 to 16 January 2021. Both OC-CCI and GlobColour Chl products were generated by bias-correcting and band-shifting specific spectral bands in the R_rs measured from several satellite sensors: Sea-viewing Wide Field-of-view Sensor (SeaWiFS), Moderate Resolution Imaging Spectroradiometer (MODIS), Medium Resolution Imaging Spectrometer (MERIS), Visible Infrared Imaging Radiometer Suite (VIIRS) and Sentinel 3A-Ocean Color and Land Imager (OLCI).

GlobColour has released two Chl products derived from the weighted averaging (AVW) and Garver, Siegel, Maritorena (GSM) algorithms, but only the AVW R_rs product is available and was used in this study. R_rs values at 443 nm, 488 nm (490 nm for OC-CCI and GlobColour) and 547 nm (560 nm for OC-CCI) from Aqua-MODIS, OC-CCI and GlobColour products were used to develop the parameter-tuned Chl retrieval algorithms. A summary of the satellite-derived Chl and R_rs products is provided in

Table 2. Summary of satellite products used in this study.

Product	Projection	Temporal Resolution	Spatial Resolution	Date
Aqua-MODIS	L3b	8 days	4 km	28 July 2002–16 January 2021
OC-CCI				29 August 1997–17 January 2021
GlobColour-AVW				29 August 1997–16 January 2021

.

2.3. Match of In Situ and Satellite Data

Co-located in situ and satellite observations were identified to evaluate and develop parameter-tuned algorithms for Chl retrieval. To ensure the quality of match-ups, the satellite image closest in time to each in situ Chl measurement was selected for matching. Then, a 3 × 3 pixel window centred on the in situ location was used to extract satellite data, following the protocol described by Bailey and Werdell [41]. The quality control procedures were as follows: (i) only match-ups containing at least 50% of valid pixels within the window were retained; (ii) outlier pixels with (pixel value – median value) > ±1.5 standard deviation were excluded; and (iii) match-ups were removed if the coefficient of variation (CV) of the remaining pixels exceeded 0.15, ensuring minimal spatial variability within the window. A full description of the match-up procedure is provided in Figure 2.

The performance of satellite-derived products was evaluated by using established statistical metrics. Bias and mean absolute difference (MAE) were employed to identify systematic over- or underestimation and quantify overall difference [38,42]. Although root mean square difference (RMSE) is most appropriate when the estimated difference distribution is Gaussian [43], it was also included in this study due to its widespread use in previous assessments of satellite-derived Chl retrievals. In addition, traditional linear regression metrics [44], including the coefficient of determination (R²), slope, and intercept between in situ Chl and satellite-derived Chl, were also recorded. The equations used to calculate bias, MAE and RMSE are presented below:

$$\mathrm{b}\mathrm{i}\mathrm{a}\mathrm{s}=10{\displaystyle \frac{{\sum }_{i=1}^n({\log }_{10}(S_i)-{\log }_{10}(O_i))}{n}}$$

(1)

$$\mathrm{M}\mathrm{A}\mathrm{E}=10{\displaystyle \frac{{\sum }_{i=1}^n\left|{\log }_{10}(S_i)-{\log }_{10}(O_i)\right|}{n}}$$

(2)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=10\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{({\log }_{10}(S_i)-{\log }_{10}(O_i))}^2}{n}}}$$

(3)

where $S_i$ and $O_i$ represent the satellite-derived Chl and in situ observation Chl, respectively. For the bias metric, values closer to unity indicate minimal systematic difference. A bias greater than unity corresponded to a positive bias (e.g., a bias of 1.10 indicates that the satellite-modelled Chl is 1.1x of the in situ observation Chl, that is, the satellite Chl represents a 10% positive bias, whereas a bias of 0.90 represents a 10% negative bias). According to Equation (2), the MAE is always positive. For example, an MAE value of 1.5 suggests that satellite Chl values are 1.5x (50%) greater on average than the in situ values.

For the regression analysis, a linear regression model was fitted between in situ observation Chl ($O_i$) and satellite-derived Chl ($S_i$):

$${\widehat{S}}_i={\beta }_0+{\beta }_1{O}_i$$

(4)

where ${\beta }_0$ and ${\beta }_1$ are the intercept and slope, respectively. The goodness of fit was assessed using the R²:

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(S_i-{\widehat{S}}_i)}^2}{{\sum }_{i=1}^n{(S_i-{\overline{S}}_i)}^2}}$$

(5)

where ${\widehat{S}}_i$ is the predicted value from the regression model, and ${\overline{S}}_i$ is the mean value of satellite-derived Chl.

2.4. Development of Parameter-Tuned Algorithm

The current Chl retrieval algorithm used by NASA for Aqua-MODIS processing is the OC3M algorithm (https://oceancolor.gsfc.nasa.gov/resources/atbd/chlor_a/; accessed on 31 May 2024). This algorithm estimates Chl concentration from the maximum band ratio (MBR) of R_rs using a single fourth-order polynomial function, expressed as:

$$\begin{gathered} {\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left(\mathrm{C}\mathrm{H}\mathrm{L}\right)=a_0+a_1\mathrm{X}+a_2{\mathrm{X}}^2+a_3{\mathrm{X}}^3+a_4{\mathrm{X}}^4 \\ \mathrm{X}={\mathrm{l}\mathrm{o}\mathrm{g}}_{10}({\displaystyle \frac{{\mathrm{m}\mathrm{a}\mathrm{x}[\mathrm{R}\mathrm{r}\mathrm{s}}_{443},{\mathrm{R}\mathrm{r}\mathrm{s}}_{488}]}{{\mathrm{R}\mathrm{r}\mathrm{s}}_{547}}}) \end{gathered}$$

(6)

where $\mathrm{C}\mathrm{H}\mathrm{L}$ is the retrieval Chl concentration (mg/m³) and $\mathrm{X}$ is the logarithm (base 10) of the MBR of R_rs. For Aqua-MODIS, the coefficients are $a_0=0.26294$, $a_1=-2.64669$, $a_2=1.28364$, $a_3=1.08209$, and $a_4=-1.76828$, as reported by O’Reilly and Werdell [45].

To develop more accurate Chl retrieval algorithms, in situ Chl were matched with corresponding satellite-derived R_rs data, following the flowchart in Figure 2. An optimization model was subsequently constructed to establish the relationship between R_rs and in situ Chl. The parameter-tuned algorithms were designed to achieve a regression slope of one, an intercept of zero, and maximize R² when comparing algorithm-derived Chl with in situ Chl. Since OC-CCI and GlobColour are merged products obtained from multi-source satellites with different retrieval algorithms, no unified and sensor-specific Chl retrieval models are associated with them. Therefore, the original OC3M algorithm was used as a baseline for initiating the optimization process across all three products (i.e., Aqua-MODIS, OC-CCI and GlobColour). Increasing and decreasing the degree of the polynomial was allowed to obtain the best fit. The performance of the optimized algorithms derived from OC-CCI and GlobColour R_rs products was evaluated in comparison to the standard OC3M algorithm.

3. Results

3.1. Initial Comparison Between Satellite Estimation and In Situ Data

Following the match-up procedure described in Section 2.3, the number of successful match-ups for each product is summarized in

Table 3. Statistical metrics for each satellite product versus in situ data.

	Aqua-MODIS	OC-CCI	GlobColour-AVW
Match-ups	1111	1329	937
Bias	0.57	0.76	0.64
MAE	2.15	1.84	1.96
RMSE	2.65	2.27	2.40
R2	0.18	0.36	0.27
Slope	0.18	0.36	0.21
Intercept	0.26	0.31	0.26

. Figure 3 represents scatter plots comparing satellite-derived Chl concentrations with in situ measurements for Aqua-MODIS, OC-CCI, and GlobColour-AVW, shown in both linear (top row) and log (bottom row) space.

All three products underestimated Chl concentrations in the SO when Chl > ~0.2/0.3 mg/m³, with bias values less than 1.00. Specifically, the bias rates were 43% for Aqua-MODIS, 24% for OC-CCI, and 36% for GlobColour-AVW. The MAE was highest for Aqua-MODIS (115%), followed by GlobColour-AVW (96%) and OC-CCI (84%). Similarly, the MRSE indicated poor retrieval accuracy, with values of 2.65 for Aqua-MODIS, 2.40 for GlobColour-AVW, and 2.27 for OC-CCI.

All products tended to overestimate Chl concentrations below ~0.2/0.3 mg/m³. Regression analysis revealed weak correlations between satellite-derived and in situ Chl, with low correlation coefficients (R² = 0.18–0.36), poor slopes (0.17–0.35) and significant offsets (intercepts = 0.26–0.31). Among the three products, OC-CCI demonstrated the best overall performance, whereas Aqua-MODIS showed the poorest agreement with in situ observations.

3.2. Design of Parameter-Tuned Algorithms

3.2.1. Refined Dataset

To support the development of parameter-optimized algorithms, match-ups were further analyzed using histograms of ${\mathrm{l}\mathrm{o}\mathrm{g}}_{10}(\mathrm{S}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{l}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{e} \mathrm{C}\mathrm{h}\mathrm{l}/\mathrm{i}\mathrm{n} \mathrm{s}\mathrm{i}\mathrm{t}\mathrm{u} \mathrm{C}\mathrm{h}\mathrm{l})$, as shown in Figure 4. The negative mode values (−0.24 mg/m³ for Aqua-MODIS, −0.12 mg/m³ for OC-CCI, and −0.20 mg/m³ for GlobColour-AVW) indicated a systematic underestimation across all products. To improve data reliability for algorithm development, only match-ups within one standard deviation of the mode were retained, thereby minimizing the influence of outliers and focusing on the core distribution. A 27% (810 of 1111), 24% (1005 of 1329), and 26% (692 of 937) reduction in match-ups was observed for Aqua-MODIS, OC-CCI and GlobColour-AVW. The statistical summary of the histograms for each satellite product is presented in

Table 4. Statistical summary of the histograms for each satellite product in Figure 4.

	Aqua-MODIS	OC-CCI	GlobColour-AVW
Match-ups (All)	1111	1329	937
Mean	−0.24	−0.12	−0.20
Standard Deviation	0.35	0.33	0.33
Match-ups (Filtered)	810	1005	692

.

3.2.2. Training and Validation Dataset

The filtered in situ Chl and R_rs match-ups (Figure 4 and Table 4) were divided into two groups: one for model training and the other for validation. To ensure the training dataset covers a wide Chl range, it includes the maximum and minimum samples from each cruise. Additionally, for each cruise, 2/3 data were randomly selected from the [minimum, median] and [median, maximum] intervals for the training data, and the remaining data were used for validation data. The numbers of training and validation samples were 535 and 275 for MODIS, 671 and 334 for OC-CCI, and 461 and 231 for GlobColour, respectively. The temporal and concentration distributions of the training (white bars) and validation (grey bars) data are shown in Figure 5.

In situ Chl datasets are distributed relatively uniformly over the 20+ years of observations. Most records were concentrated in 2004, 2006, 2008, 2011, 2012, 2016, and 2017, mainly from the Seabass, Palmer, and AWI cruises. The concentration distribution showed that the distribution is approximately equal between training and validation datasets. The training dataset covered a Chl range of 0.061–7.490 mg/m³ for Aqua-MODIS, 0.049–18.627 mg/m³ for OC-CCI, and 0.049–6.527 mg/m³ for GlobColour. The validation dataset also encompassed a wide Chl range, supporting comprehensive model evaluation.

3.2.3. Parameter-Tuned Algorithms

Using the training dataset and algorithm development procedure described in Section 2.4, fourth-order OCx polynomial functions were derived for Aqua-MODIS, OC-CCI and GlobColour-AVW, respectively. Figure 6 shows the polynomials relating in situ Chl to MBR, and the corresponding optimized polynomial equations are listed below. For the GlobColour-AVW dataset, the fourth-order polynomial shows an unrealistic, non-physical behaviour for MBRs larger than about 5. In this case, a more realistic model is obtained with a third-order polynomial function (see Figure 6c).

Aqua-MODIS:

$$\begin{gathered} {\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left({\mathrm{C}\mathrm{H}\mathrm{L}}_{\mathrm{A}\mathrm{M}}\right)=0.41812-2.39956{\mathrm{X}}_{\mathrm{A}\mathrm{M}}+2.58460{{\mathrm{X}}_{\mathrm{A}\mathrm{M}}}^2-5.018051{{\mathrm{X}}_{\mathrm{A}\mathrm{M}}}^3+2.96640{{\mathrm{X}}_{\mathrm{A}\mathrm{M}}}^4 \\ {\mathrm{X}}_{\mathrm{A}\mathrm{M}}={\mathrm{l}\mathrm{o}\mathrm{g}}_{10}({\displaystyle \frac{{\mathrm{m}\mathrm{a}\mathrm{x}[\mathrm{R}\mathrm{r}\mathrm{s}}_{443},{\mathrm{R}\mathrm{r}\mathrm{s}}_{488}]}{{\mathrm{R}\mathrm{r}\mathrm{s}}_{547}}}) \end{gathered}$$

(7)

OC-CCI:

$$\begin{gathered} {\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left({\mathrm{C}\mathrm{H}\mathrm{L}}_{\mathrm{O}\mathrm{C}}\right)=0.33370-1.87432{\mathrm{X}}_{\mathrm{O}\mathrm{C}}+1.61855{{\mathrm{X}}_{\mathrm{O}\mathrm{C}}}^2-1.51217{{\mathrm{X}}_{\mathrm{O}\mathrm{C}}}^3-0.38774{{\mathrm{X}}_{\mathrm{O}\mathrm{C}}}^4 \\ {\mathrm{X}}_{\mathrm{O}\mathrm{C}}={\mathrm{l}\mathrm{o}\mathrm{g}}_{10}({\displaystyle \frac{{\mathrm{m}\mathrm{a}\mathrm{x}[\mathrm{R}\mathrm{r}\mathrm{s}}_{443},{\mathrm{R}\mathrm{r}\mathrm{s}}_{490}]}{{\mathrm{R}\mathrm{r}\mathrm{s}}_{560}}}) \end{gathered}$$

(8)

GlobColour-AVW:

$$\begin{gathered} {\mathrm{l}\mathrm{o}\mathrm{g}}_{10}\left({\mathrm{C}\mathrm{H}\mathrm{L}}_{\mathrm{G}\mathrm{C}}\right)=0.39223-2.98077{\mathrm{X}}_{\mathrm{G}\mathrm{C}}+5.39321{{\mathrm{X}}_{\mathrm{G}\mathrm{C}}}^2-9.65322{{\mathrm{X}}_{\mathrm{G}\mathrm{C}}}^3+5.79302{{\mathrm{X}}_{\mathrm{G}\mathrm{C}}}^4 \\ {\mathrm{X}}_{\mathrm{G}\mathrm{C}}={\mathrm{l}\mathrm{o}\mathrm{g}}_{10}({\displaystyle \frac{{\mathrm{m}\mathrm{a}\mathrm{x}[\mathrm{R}\mathrm{r}\mathrm{s}}_{443},{\mathrm{R}\mathrm{r}\mathrm{s}}_{490}]}{{\mathrm{R}\mathrm{r}\mathrm{s}}_{547}}}) \end{gathered}$$

(9)

The optimized algorithms effectively reduced the underestimation of Chl concentrations, as shown in Figure 7 and

Table 5. Accuracy statistics for both the original OC3M algorithms and the new parameter-tuned algorithms to the training data in the Southern Ocean. Dynamic values represent the change in each metric relative to its reference value (i.e., one for bias, R², and slope; zero for MAE, RMSE, and intercept), indicating improvements after parameter tuning.

	Aqua-MODIS			OC-CCI			GlobColour-AVW
	OC3M	New	Dynamic	OC3M	New	Dynamic	OC3M	New	Dynamic
Bias	0.68	1.00	0.32	0.57	1.00	0.43	0.72	1.00	0.28
MAE	1.59	1.39	0.20	1.81	1.37	0.44	1.52	1.37	0.15
MRSE	1.78	1.50	0.28	2.01	1.47	0.54	1.67	1.48	0.19
R2	0.70	0.69	0.01	0.78	0.79	0.01	0.77	0.67	0.10
Slope	0.45	0.67	0.22	0.79	0.72	0.07	0.49	0.77	0.28
Intercept	0.10	0.17	0.07	−0.09	0.18	0.27	0.11	0.11	0.00

. Bias was reduced by 32% for Aqua-MODIS, 43% for OC-CCI and 28% for GlobColour-AVW. Corresponding reductions in MAE reached 20%, 44%, and 15%, and RMSE decreased by 28%, 54%, and 19%, respectively. Slope estimates improved notably for Aqua-MODIS (from 0.45 to 0.67) and GlobColour-AVW (from 0.49 to 0.77), whereas OC-CCI showed a slight decrease (from 0.79 to 0.72). R² remained relatively stable for Aqua-MODIS (from 0.70 to 0.69) and OC-CCI (from 0.78 to 0.79), but decreased for GlobColour-AVW (from 0.77 to 0.67). Intercept values for Aqua-MODIS and GlobColour-AVW exhibited minimal changes (< 0.10).

3.3. Independent Evaluation

To further evaluate the robustness of parameter-tuned algorithms, satellite-derived Chl concentrations were recalculated using the optimized algorithms and compared to the validation dataset (Figure 8 and

Table 6. As in Table 5, but for the validation data.

	Aqua-MODIS			OC-CCI			GlobColour-AVW
	OC3M	New	Dynamic	OC3M	New	Dynamic	OC3M	New	Dynamic
Bias	0.68	1.02	0.30	0.57	0.99	0.42	0.70	0.97	0.27
MAE	1.58	1.42	0.16	1.84	1.39	0.45	1.56	1.38	0.18
MRSE	1.77	1.51	0.26	2.07	1.49	0.58	1.73	1.50	0.23
R2	0.69	0.68	0.01	0.81	0.84	0.03	0.81	0.91	0.10
Slope	0.42	0.62	0.20	1.02	0.92	0.10	0.35	0.76	0.41
Intercept	0.12	0.21	0.09	−0.26	0.03	0.29	0.19	0.10	0.09

). Consistent with the results in Section 3.2.3, the optimized algorithms substantially reduced Chl underestimation. Bias decreased by 30% for Aqua-MODIS, 42% for OC-CCI and 27% for GlobColour-AVW. Corresponding reductions in MAE were 16%, 45% and 18%, while RMSE declined by 0.26 (from 1.77 to 1.51), 0.58 (from 2.07 to 1.49) and 0.23 (from 1.73 to 1.50), respectively. Slope values improved notably for Aqua-MODIS (0.20) and GlobColour-AVW (0.41), but decreased slightly for OC-CCI (0.10). R² showed a slight improvement for OC-CCI (0.03) and a notable increase for GlobColour-AVW (0.10), while a slight decrease of 0.01 was observed for Aqua-MODIS. Additionally, the intercept for OC-CCI (0.03) and GlobColour-AVW (0.10) shifted closer to zero compared to original algorithms.

4. Discussion

Accurate satellite-derived Chl estimates are essential for detecting long-term and large-scale phytoplankton dynamics. However, existing retrieval algorithms have been shown to underestimate Chl concentrations in the SO [19,21,24,25]. This study evaluated the application performance of Aqua-MODIS, OC-CCI, and GlobColour-AVW products in the SO and developed parameter-tuned Chl algorithms to improve the Chl retrieval accuracy.

Among the three satellite products evaluated, OC-CCI exhibited the best overall performance (R² = 0.36, Slope = 0.36), followed by GlobColour-AVW (R² = 0.27, Slope = 0.21), whereas Aqua-MODIS showed the poorest agreement (R² = 0.18, Slope = 0.18) with in situ observations (Figure 3 and Table 3). These results suggest that multi-mission merged products (i.e., OC-CCI and GlobColour-AVW) perform better in the SO than single-sensor datasets, aligning with previous findings [46,47].

Despite OC-CCI’s comparatively better performance, all products systematically underestimated Chl concentrations in the SO. The average underestimations were 43%, 24%, and 36% for Aqua-MODIS, OC-CCI, and GlobColour-AVW, respectively, with the largest biases observed at higher Chl concentrations (>0.2 mg/m³ for Aqua-MODIS and GlobColour-AVW; >0.3 mg/m³ for OC-CCI). A regionally tailored Chl algorithm is essential to correct this underestimation.

To address this, we combined high-quality (HPLC-derived), long-term (1997–2021) and spatially widespread (south of 40°S) in situ measurements to develop robust and reliable algorithms. The new parameter-tuned empirical algorithms reduced the systematic bias to 1% (OC-CCI), 3% (GlobColour-AVW), and a slight overestimation of 2% for Aqua-MODIS (Figure 8 and Table 6). All products showed marked improvements in performance, with R² increasing to 0.68–0.91, slopes approaching 1.0 (0.62–0.92), and notable reductions in MAE (1.39–1.42) and RMSE (1.49–1.51). These results compare favourably with previous studies [24,28], demonstrating the effectiveness of the parameter-tuned approach.

Nevertheless, several limitations warrant consideration. The parameter-tuned algorithms are SO-specific and are not directly applicable to other oceanic regions (e.g., the global oceans). The uneven spatiotemporal distribution of in situ data and the potential impact of climate-driven shifts in ocean optical properties may limit the long-term reliability of the optimized algorithm. Regular updates incorporating newly collected in situ observations are therefore essential for maintaining algorithm accuracy in this climatically sensitive region. In addition, only the OC3M empirical algorithm was considered in this study.

This study uses satellite-derived R_rs values, which are affected by atmospheric correction errors. In some way, this means that the algorithms are taking these errors into account. Collecting larger datasets of field radiometry measurements from which R_rs can be derived should, however, be a priority in the effort to improve the accuracy of satellite ocean colour products in the SO. Such datasets are still extremely scarce for the SO.

Future work should also explore semi-analytical algorithms [48] and data-driven methods such as machine learning [49], which may better capture the region’s complex bio-optical variability. Moreover, the large solar zenith angle observation conditions [50], complex bio-optical properties [23], and sea ice [21] potentially affect the retrieval accuracy of Chl algorithms. Future research is needed to systematically evaluate potential drivers on Chl estimation in the SO.

5. Conclusions

This study assessed the performance of three satellite-derived Chl products (i.e., Aqua-MODIS, OC-CCI, and GlobColour-AVW) in the SO, and developed parameter-tuned empirical algorithms based on high-quality, long-term in situ Chl measurements. The multi-mission OC-CCI dataset showed the best agreement with in situ observations, followed by GlobColour-AVW, while Aqua-MODIS performed the poorest. All products systematically underestimated Chl concentrations, particularly at high biomass levels. The newly developed algorithms significantly improved retrieval accuracy, reducing biases to within 3%, increasing R² (0.68–0.91), and decreasing MAE and RMSE.

Despite the improved performance, limitations remain due to the uneven spatiotemporal distribution of in situ data and potential changes in optical water properties driven by climate variability. Continued algorithm refinement, supported by regular updates with new in situ observations, is essential for reliable long-term monitoring, as well as increasing data coverage for the ocean reflectance. Future work should also incorporate semi-analytical and machine learning approaches to better address the SO’s complex bio-optical characteristics and challenging observation conditions.