A Novel Methodology to Correct Chlorophyll-a Concentrations from Satellite Data and Assess Credible Phenological Patterns

Abstract

Remote sensing data play a crucial role in capturing and evaluating eutrophication, providing a comprehensive view of spatial and temporal variations in water quality parameters. Chlorophyll-a concentration time series analysis aids in understanding the current trophic state of coastal waters and tracking changes over time, enabling the evaluation of water bodies’ trophic status. This research presents a novel and replicable methodology able to derive accurate phenological patterns using remote sensing data. The methodology proposed uses the two-decade MODIS-Aqua surface reflectance dataset, analyzing data from 30-point stations and calculating chlorophyll-a concentrations from NASA’s Ocean Color algorithm. Then, a correction process is implemented through a robust, simple statistical analysis by applying LOESS smoothing to detect and remove outliers from the extensive dataset. Different scenarios are reviewed and compared with field data to calibrate the proposed methodology accurately. The results demonstrate the methodology’s capacity to produce consistent chlorophyll-a time series and to present phenological patterns that can effectively identify key indicators and trends, resulting in valuable insights into the coastal body’s trophic state. The case study of the Ambracian Gulf is characterized as hypertrophic since algal bloom during August reaches up to 5 mg/m³, while the replicate case study of Aitoliko shows algal bloom reaching up to 2.5 mg/m³. Finally, the proposed methodology successfully identifies the positive chlorophyll-a climate tendencies of the two selected Greek water bodies. This study highlights the value of integrating statistical methods with remote sensing data for accurate, long-term monitoring of water quality in aquatic ecosystems.

1. Introduction

Eutrophication of coastal and inland water bodies is induced by the extensive enrichment of nutrients [1] fueled by human activities, such as agricultural runoff, untreated wastewater, and industrial discharges [2], leading to overgrowth of algae and a decline in water quality [3]. Coastal ecosystems are among the most productive and biodiverse habitats on the planet today, providing critical services, such as food production, coastal defense, and nutrient cycling, along with recreation, etc. [4]. Therefore, algae blooms, induced by eutrophication, can release pollutants that are harmful to both aquatic life and human health. Furthermore, parallel to water quality deterioration, eutrophication reduces water oxygen levels, potentially resulting in anoxic “dead zones,” which do not allow the survival of the aquatic ecosystem.

When addressing eutrophication, different indicators exist to quantify the trophic status change, such as chlorophyll-a [5,6], nutrient concentrations [7], water clarity/transparency [8,9], and dissolved oxygen [10,11]. The seasonal distribution of chlorophyll-a stands out as a significant factor for evaluating water bodies’ trophic status [12,13] through remote sensing applications. In remote sensing, chlorophyll-a often serves as a predominant indicator due to its direct measurability and the existence of well-established algorithms for extracting it from satellite data. The phenological patterns of chlorophyll-a offer important insights into tracking the spatial and temporal changes in eutrophication [14], which help in understanding how ecosystems respond to both natural changes and human impacts [15,16].

Remote sensing is a powerful tool for water quality management, enabling large-scale monitoring of key parameters such as chlorophyll-a, turbidity, and dissolved organic matter. It utilizes satellite sensors (e.g., MODIS, Sentinel-2, Landsat) for broad spatial coverage, airborne sensors for high-resolution data, and drone-mounted or in situ hyperspectral sensors for detailed local assessments. These technologies provide real-time and historical insights, supporting early detection of pollution, algal blooms, and ecosystem changes, ultimately aiding in sustainable water quality management.

In this effort, long-term remote sensing datasets capture variations in chlorophyll-a concentrations [17], timing, duration, and intensity on both seasonal and interannual timescales; reveal shifts in chlorophyll-a phenological patterns; and reflect trophic status variability, which is closely tied to climate change [18]. Additionally, chlorophyll-a phenological patterns can serve as early indicators of harmful algal blooms (HABs), which pose significant ecological and socio-economic challenges [19]. These insights are essential for effective fisheries management, conservation strategies, and monitoring water quality [20].

MODIS Aqua and SeaWiFS satellite datasets on chlorophyll-a processed into Level 3 (L3) products are readily accessible and have been heavily utilized in coastal studies and administration [21,22]. Nevertheless, researchers express hesitation about the precision of these L3 products [21,23,24] in depicting the true phenological trends of coastal chlorophyll-a, particularly in coastal water bodies. Owing to this growing interest in chlorophyll-a assessment, various machine learning models, including Random Forest [25,26,27], principal component analysis [12], and many other statistical methodologies, have been introduced in an effort to improve the reliability of chlorophyll-a datasets. However, these models typically require extensive training datasets and can introduce complexity in interpretation. Given the study’s focus on long-term phenological analysis with limited available field data for calibration, simpler statistical approaches are often preferable and a more practical choice. Current L3 chlorophyll-a datasets, being spatially and temporally averaged, often smooth out variations excessively, thereby masking crucial ecological signals. Research indicates that L3 data inadequately capture the genuine variability of algae blooms, especially in diverse coastal ecosystems with shallow waters [28,29]. The discrepancies in these datasets emphasize the necessity of introducing a methodology that improves satellite algorithms for accurate phenological analyses.

This study aspires to introduce a simple, yet novel, methodology to generate accurate chlorophyll-a patterns refining L2-derived chlorophyll-a satellite datasets, enabling precise identification of phenological patterns. Through the integration of MODIS Aqua data with advanced breakpoint detection, the study assesses the potential for enhanced temporal and spatial resolutions of coastal water bodies. The Ambracian Gulf serves as the primary case study and the Aitoliko Lagoon as a replicate study, demonstrating how improved processing can reveal true phenological patterns and trophic status dynamics. This reproducible method strives to bridge existing gaps in satellite data processing, furnishing practical insights for the monitoring and management of aquatic ecosystems.

2. Materials and Methods

2.1. Satellite Data

MODIS Aqua surface reflectance data (MYDOCGA) were acquired using the RStudio (version 4.3.2) programming environment and the R programming language. The MODIS Aqua data were selected due to their long-term dataset availability (from 2002 until today) and their ability to detect the phenological patterns of water bodies’ quality characteristics [16]. The dataset was downloaded from NASA’s Ocean Color Web portal (https://oceancolor.gsfc.nasa.gov, accessed on 22 January 2025) and spanned from 2002 to 2023, with a 1 km spatial resolution and daily temporal coverage. The MYDOCGA product is a Level 2 ocean reflectance product and includes 9 spectral channels, B8 through B16, which are crucial for estimating chlorophyll-a concentrations for water bodies using ocean color algorithms. For this study, a total of 7502 georeferenced hdf images were downloaded and extracted for point stations within the Area of Interest (AOI).

In addition, MODIS-Aqua ready-to-use chlorophyll-a concentrations were acquired to prove that the deviation of satellite-derived chlorophyll-a seasonality is due to air interference, sensor limitations, and the complexity of coastal and inland water bodies [30]. In total, 217 Level 3 MODIS Aqua chlorophyll-a concentrations were downloaded for the Area of Interest from NASA Ocean Color Web Portal, with a spatial resolution of 4 km and monthly temporal coverage.

2.2. Case Study: Ambracian Gulf

The Ambracian Gulf, located in the region of Epirus, Greece, was chosen as a case study to validate the proposed methodology, due to its ecological relevance and well-documented eutrophication record. The Ambracian Gulf is known for its diverse ecosystem and rich biodiversity. It covers an area of about 405 km² and has an average depth of 35 m [31]. The gulf is connected to the Ionian Sea through the Aktio–Preveza channel and is a typical example of a semi-enclosed bay, of fjord type, in the Mediterranean, since its only point of communication with the open sea is this narrow channel [32]. The gulf’s semi-enclosed geomorphology causes restricted water exchange with the Ionian Sea, leaving it especially vulnerable to nitrogen loading from agricultural runoff and urban effluents.

Efforts have been made to improve and maintain water quality in the Ambracian Gulf. The water quality of the Ambracian Gulf is characterized as highly eutrophic, taking into account the measurements of orthophosphate ions of the surface layer, while using the data of nitrate and ammonia nitrogen, as well as chlorophyll a, the water status of the gulf is characterized as eutrophic [32].

2.3. Field Data During 2010–2011

During the period 2010–2011, physicochemical measurements were carried out in the Ambracian Gulf to validate satellite-derived chlorophyll-a concentrations and calibrate the proposed methodology. A total of six surveys were conducted over nine months, from September 2010 to July 2011. These cruises covered three sampling locations within the Ambracian Gulf, with each station being visited at two-month intervals. Surface sampling was carried out at stations numbered 6, 14, and 26, as seen in Figure 1. The water bottle samples were placed in black bags to avoid light exposure, preserved in a portable fridge, and transferred to the laboratory within two hours after collection.

Chlorophyll-a measurements were conducted using 90% ethanol and measured with a UV2401 Perkin Elmer spectrophotometer [33]. Chlorophyll-a concentration was calculated using the trichromatic spectrophotometry method, as described by the equations of Jeffrey and Humphrey (1975) [34]:

$$Chl\text{-}a \left({\displaystyle \frac{\mathrm{mg}}{{\mathrm{m}}^3}}\right)=\left(11.85E_{664}-1.54E_{647}-0.08E_{630}\right)\times {\displaystyle \frac{V_{acetone}}{V_{water}}}$$

(1)

where E₆₆₄, E₆₄₇, and E₆₃₀ are, respectively, the absorptions at 664, 647, and 630 nm (after subtracting the absorption of E₇₅₀), V_water is the volume of the water sample, and V_acetone is the extracted volume of acetone (mL).

2.4. Pre-Processing Analysis

Chlorophyll-a datasets were derived employing the NASA Ocean Color algorithm. Raw MODIS Aqua surface reflectance data underwent a series of pre-processing steps to ensure the accuracy and reliability of the estimated values of the chlorophyll-a datasets. The MYDOCGA product used consisted of Level-2 processed georeferenced images, which were corrected for atmospheric conditions, including gases, aerosols, as well as Rayleigh scattering. Surface reflectance data were further extracted to match the spatial (1 km) and temporal (daily) resolutions of the 30 satellite monitoring stations closed within the Area of Interest.

Then, chlorophyll-a concentrations were calculated using NASA’s Ocean Color algorithm, which has been extensively validated for general use in determining the (biological and chemical) characteristics of coastal and open ocean waters [22,23,35]. This algorithm employs a four-band ratio approach. For the MYDOCGA product, the spectral channels used for the chlorophyll-a algorithm were B9 (443 nm), B10 (490 nm), B12 (555 nm), and B14 (670 nm). The resulting concentrations of chlorophyll-a were retrieved for each observation station as a function of time.

To ensure data quality and reduce the impact of anomalies of the L2 dataset, the pre-processing analysis also included filtering out invalid pixels based on quality flags to resolve data gaps and outliers caused by cloud cover, sensor limitations, and other abnormalities.

The first logical control introduced suggests that the chlorophyll-a values cannot differ for two consecutive days above 25% for the same point stations within the study area. This control was often used to detect cloudy days within the dataset and was dominant in multiple point stations on the same day for the study area. The second logical control introduced aimed to restrict chlorophyll-a values within a logical chlorophyll-a concentration range, and after calibration with field data, 50 mg/m³ was identified as the best fitting range. This step ensured the quality of the data by reducing the effect of incorrect values caused by malfunctioning sensors, cloud interference, or many other anomalies [36,37]. In this study, non-equivalent data were attributed as missing values, and they were not removed from the initial dataset.

2.5. Breakpoint Processing and LOESS Smoothing

The Locally Estimated Scatterplot Smoothing (LOESS) algorithm [38,39] was applied to identify breakpoints in the time series data, which indicate significant changes in chlorophyll-a concentrations such as algal blooms, shifts in nutrient availability, or environmental disturbances that influence chlorophyll-a dynamics and provide insights into long-term ecological processes. The timing and number of breakpoints were determined by performing forward and reverse correction, computing the relative change between consecutive values (or in reverse order). Observations with relative changes greater than 1.25 were removed, with additional filtering applied iteratively to ensure consistency. This selection ensured that chlorophyll-a concentration changes in water bodies greater than 25% are unlikely to occur in a time period of two consecutive days. Then, a non-parametric regression analysis (LOESS smoothing) was applied to capture temporal trends in the measurements while accounting for local dependencies.

The application of the LOESS algorithm aimed to fulfill two primary purposes: first, to complete the missing values in the pre-processed dataset, and second, to correct the chlorophyll-a estimates by capturing local variations and smoothing out noise and outliers in the satellite data, resulting in a more accurate portrayal of genuine chlorophyll-a concentrations. LOESS is a non-parametric regression method that applies localized polynomial models to segments of data within a moving window [40]. Therefore, LOESS describes trends by fitting a local polynomial around each data point, capturing localized changes in the data rather than the overall trend, as it effectively captures moderate nonlinearities [41] in the dynamics of chlorophyll-a. Moreover, while traditional regression methods typically estimate the conditional mean of a response variable, LOESS smoothing captures localized patterns and moderate non linearities in data, making it more robust to noise and fluctuations in time series dynamics [42].

The LOESS algorithm estimates a smooth functional relationship $f$ between the predictor and response variables, specifically the chlorophyll-a estimates $y_i$ and the time $t_i,$ respectively, such that

$$y_i=f\left(t_i\right)+{ϵ}_i$$

(2)

where ${ϵ}_i$ is an error term, which is assumed to have a mean of zero. An important parameter of the LOESS algorithm is the span parameter, $\alpha ,$ which defines a local neighborhood for every $t_i$ by selecting its $k=\lfloor \alpha n\rfloor$ nearest-neighbor points, where $\lfloor \dot{}\rfloor$ is the floor function and $n$ the sample size to fit a local polynomial by assigning a weight to each data point based on its distance from $t_i$. A small span in LOESS results in a highly flexible fit that closely follows local variations but may introduce noise and overfitting, while a large span produces a smoother, more stable trend by averaging over a broader range of data, potentially oversmoothing important local patterns.

The span size for LOESS was 0.03 through cross-validation on independent station datasets to reduce residual errors. It should be noted that the span parameter is a positive real number specifying the size of the data neighborhood, and since it is a fraction of the total data, no unit is applied. This 0.03 span provided an optimal balance between trend adherence and noise reduction. The proposed methodology introduced backward and forward chlorophyll-a concentration correction to each point station of the long time series.

Previous studies have demonstrated the effectiveness of LOESS in various fields beyond remote sensing [43]. In remote sensing applications, LOESS has been successfully employed to correct inaccuracies caused by minor contaminations in water area mapping derived from Landsat imagery, achieving a relative root-mean-squared error of 2.2% [44]. Additionally, LOESS-based cloud-filling techniques can diminish errors in the estimation of regional mean chlorophyll-a concentrations by 50% to 80%, especially in regions characterized by significant cloud cover [45].

The application of LOESS smoothing in the estimation of chlorophyll-a from satellite data can greatly enhance the reliability of the resultant products, thereby empowering researchers to make better-informed decisions concerning the management of aquatic ecosystems [46,47].

2.6. Validation and Versatility of the Methodology Proposed

To validate the proposed methodology and review the derived phenological patterns compared to Level-3 ready-to-use chlorophyll-a datasets, the analysis was replicated in another coastal water body, the Aitoliko Lagoon (Figure 2). This site was chosen for its distinct hydrological characteristics and susceptibility to algal blooms, providing a contrasting environment to the Ambracian Gulf for testing the methodology’s robustness, adaptability, and effectiveness in capturing phenological patterns in diverse aquatic ecosystems. For this case study, the Level-2 (L2G) MYDOCGA surface reflectance Modis Aqua image dataset was utilized along with the Level-3 (L3) chlorophyll-a concentrations Modis Aqua dataset from NASA Ocean Color. Within the Aitoliko Lagoon, 18 pixel point stations were identified for the L2G dataset with a 1 km spatial resolution and daily temporal coverage and 1 point station for the L3 dataset with a 4 km spatial resolution and monthly temporal coverage.

2.7. Temporal Coverage of Satellite Data

2.7.1. Monthly Temporal Coverage

Mean monthly temporal coverage of the long daily time series dataset helps reduce data volume for better time series review and improved consistency [1] capturing, while capturing regular variations and possible trends.

The formula for calculating the mean monthly composite for a given pixel can be expressed as

$$M{M}_p={\displaystyle \frac{1}{n_v}} {\sum }_{i=1}^{n_d}{D}_p,i$$

(3)

where

$M{M}_p$ is the mean monthly composite for point-pixel “p”;

$n_v$ is the number of “valid”, non-missing, daily observations for pixel “p” within the month;

$n_d$ is the total number of days in the month; and

$D_{p,i}$ is the daily observation value for point-pixel “p” on day “i” (missing values are excluded from the average calculation within the month).

2.7.2. Seasonal Temporal Coverage

Mean seasonal temporal coverage offers valuable insights for assessing and reviewing water bodies’ phenological patterns. Seasonal composites are widely used in remote sensing to monitor marine ecosystem health and productivity [48], to assess water quality and pollution impacts [49,50], to understand climate change impacts [17,51,52], to evaluate fisheries management and aquaculture [53], and to develop ocean color remote sensing algorithms [49] as well.

The formula for calculating the mean seasonal composite for a given pixel can be expressed as

$$M{S}_p={\displaystyle \frac{1}{n_m}} {\sum }_{m=1}^{n_m}M{M}_p$$

(4)

where

$M{S}_p$ is the mean seasonal composite for point-pixel “p”;

$n_m$ is the number of “months” of observation for pixel “p”; and

$M{M}_p$ is the mean monthly composite for point-pixel “p” in month “m”.

2.7.3. Yearly Temporal Coverage

Mean yearly temporal coverage offers valuable insights for assessing long-term trend analysis, climate change impacts [54], interannual variability [54], and the effectiveness of management strategies [54].

The formula for calculating the mean yearly composite for a given pixel can be expressed as

$$M{Y}_p={\displaystyle \frac{1}{12}} {\sum }_{y=1}^{12}M{M}_p$$

(5)

where

$M{Y}_p$ is the mean yearly composite for point-pixel “p” and

$M{M}_p$ is the mean monthly composite for point-pixel “p”.

3. Results

3.1. LOESS Model Calibration Procedure with Field Measurements

Field data measurements were compared to different algorithm scenarios to determine the optimal parameters for the pre-processing cloud-range limit, as described in Section 2.4, and the span of the LOESS fitting, as described in Section 2.5. The scenarios evaluated included different maximum chlorophyll-a concentration thresholds and LOESS span parameters.

To determine the criteria of accepted chlorophyll-a range concentration, the selection of the upper limit attributed to outliers included different scenarios of chlorophyll-a maximum value. The selected maximum values tested were 10, 50, 100, and 1000 mg/L. Setting 1 includes observations with values smaller than or equal to the maximum chlorophyll-a value of 10, Setting 2 includes observations with values smaller than or equal to the maximum chlorophyll-a value of 50, and Setting 3 includes observations with values smaller than or equal to the maximum chlorophyll-a value of 100.

Considering the LOESS span values, different scenarios were evaluated within the range 0.02 to 0.1, aiming to obtain the local patterns and avoid smoothing the dataset. The selected span value was 0.03, since it was the smallest span value that resulted in a dataset of complete phenological cycles per year. In addition, lower LOESS spans yield more smoothing, potentially eliminating short-term variations and hiding the dataset’s phenological patterns. The aim of the methodology described was to obtain a complete chlorophyll-a dataset that does not eliminate peak values. All the three scenarios are summarized in

Table 1. Comparison of chlorophyll-a concentration scenarios derived from the proposed methodology with field data measurements and existing satellite products. The table includes maximum chlorophyll-a thresholds, LOESS span values, line colors corresponding to Figure 2, and scenario descriptions. The N/A indicates that the added process control is non-applicable (N/A).

Scenario	Max Chlorophyll-a (mg/m3)	LOESS Span	Line Color	Description
A1	10	0.03	Green	Conservative threshold with light smoothing
A2	50	0.03	Magenta	Moderate threshold with light smoothing
A3	100	0.03	Purple	High threshold with light smoothing
Initial	N/A	N/A	Black	Very high threshold with small smoothing
L3	N/A	N/A	Red	L3 chlorophyll-a product
Field	N/A	N/A	Orange	Field data measurements for validation

. In the same table, field data measurements and existing satellite products are also given for evaluation and reference.

The application of LOESS is limited by the requirement for a large dataset and the potential impact of outliers [39]. However, the extensive time series provided by the daily surface reflectance dataset, covering approximately 20 years, helps address these concerns by offering sufficient data points for effective model fitting. In addition, the range limit introduced for cloud presence mitigates errors that could arise from this nonlinear formula of the LOESS algorithm.

Figure 3 presents the compared results of monthly chlorophyll-a values between different scenarios of maximum chlorophyll-a concentrations and the span number of the LOESS smoothing, with the field data measurements. Monthly chlorophyll-a composites were selected to present and compare the L3 ready-to-use monthly chlorophyll-a values with those produced by the methodology. Therefore, the high-accuracy field measurements of chlorophyll-a are considered representative values for the whole month in question.

The results indicate that scenario A2, with a maximum chlorophyll-a value of 50 and a LOESS span of 0.03 (magenta), best correlates with the field measurements due to the amplitude and timing of the detected peaks. The comparison highlights that scenario A3, with elevated maximum chlorophyll-a values (100, purple line), overestimates chlorophyll-a concentrations, particularly during intervals of diminished chlorophyll-a levels. The selection of higher chlorophyll-a concentrations would affect the dataset’s scaling and normalization. Assigning a maximum chlorophyll-a value of 50, we can provide an appropriate equilibrium between overestimation and underestimation relative to field observations.

The LOESS smoothing span parameter is essential for balancing adherence to the underlying data with the attenuation of noise. A duration of 0.03 offers enhanced temporal resolution, guaranteeing precise capturing of temporal patterns without excessive smoothing. The comparison with field measurements examines the essential relationship between parameter selection and methodological efficacy in estimating chlorophyll-a concentration. The findings indicate that the suggested approach, when set with a maximum chlorophyll-a concentration of 50 and a LOESS span of 0.03, provides the most precise depiction of field measurements.

3.2. Long-Term Chlorophyll-a Phenological Patterns

A continuous 20-year time series of chlorophyll-a concentrations was produced using the above methodology for the 30 monitoring sites spread throughout the Ambracian Gulf. The time series showed a clear seasonal trend, with spring and fall exhibitingthe highest values of chlorophyll-a, which corresponded to the periods when phytoplankton blooms occurred. The mean concentrations of chlorophyll-a averaged for the polygon area of the Ambracian Gulf was 4.80 mg/m³, whereas the minimum and maximum monthly reported value for the whole Ambracian Gulf ranged from 0.26 to 15.12 mg/m³, after reviewing chlorophyll-a values from July 2002 to February 2023.

To evaluate the performance of the LOESS correction, both qualitative and quantitative evaluations were carried out. The pre-processing methodology of the case study of the Ambracian Gulf resulted in 103,810 missing values out of 225,060 chlorophyll-a values generated after 7502 days of collection for 30 monitoring stations. In total, 46% of the dataset was reported as “NA” after the LOESS procedure.

Figure 4 presents the daily distribution of chlorophyll-a and the trends for point station 26 of the Ambracian Gulf after the LOESS procedure proposed in this study. Point 26 is the deepest point station of the dataset and best represents the temporal distribution of the semi-closed water body of the Ambracian Gulf [12]. Figure 4 describes significant variation in the raw data (black dots) throughout the observation period, with values mostly falling between 0 and 20 mg/m³, while there are sporadic peaks that approach 100 mg/m³. Periodic increases in phytoplankton biomass are highlighted by this variability, which is probably caused by seasonal or environmental factors like temperature, hydrodynamic conditions, or nutrient availability. The red line depicting the LOESS-produced dataset shows a period pattern with repeated peaks and ventricles. Finally, the blue shaded ribbon shows the trend’s confidence interval, with a standard error of ±1.96 around the smoothed line. The width of the confidence interval fluctuates during the 20-year time period, which is relevant to the measurement dispersion/fluctuation.

The monthly composites of the polygon area of Ambracian Gulf after the pre-processing of L2G data and the processed data with the LOESS methodology throughout the study period (2002–2023) are described in Figure 5. The L3 ready-to-use chlorophyll-a product (blue regression line: $12.378-0.05378(year-2002)$; p-value for the slope: 0.66) exhibits increased seasonal peaks, with concentrations often above 60 mg/m³. These errors introduced in the L3 values, as demonstrated in prior research [21], are likely attributable to sensor limitations and fundamental processing assumptions that neglect the intricacies of coastal and inland waterways. The L2G SR product (black regression line: $8.302+0.05165(year-2002)$; p-value for the slope: 0.183) has higher values, following chlorophyll-a true values, but also exaggerates chlorophyll-a concentrations, especially during peak development phases.

The suggested methodology (red regression line: $3.025+0.05692(year-2002)$, p-value for the slope: 0.005) effectively diminishes the overestimation of chlorophyll-a concentrations while not smoothing the variability of the chlorophyll-a concentrations, resulting in outcomes that align more closely with field data. This is especially apparent in the consistency and reduced magnitude of the seasonal peaks, which correspond more closely to the observed phenological patterns of the Ambracian Gulf. The seasonal peaks obtained by the suggested technique result in a chlorophyll-a maximum of 5 mg/m³, providing a more accurate depiction of chlorophyll-a concentrations in this intricate aquatic system. In addition, the linear fitted curve of the datasets reveals that chlorophyll-a concentrations resulting both from the L2G SR product and the proposed methodology successfully present a positive trend (statistically significant for the proposed methodology) of chlorophyll-a for the period of analysis, from 2002 to the end of 2023, whereas the L3 dataset misrepresents a negative trendline.

3.3. Seasonal Patterns of the Ambracian Gulf

Figure 6 illustrates the comparison of chlorophyll-a concentrations (mg/m³) obtained from three datasets—Level-3 MODIS Aqua chlorophyll-a (L3 chl-a product), Level-2 surface reflectance (L2G SR product), and the proposed methodology—to show the seasonal composites and evaluate phenological patterns. The L3 chl-a product (blue line) exhibits the greatest concentrations, with significant seasonal peaks beyond 20 mg/m³ in the spring and late fall. Nonetheless, these values are probably caused by the algorithm outliers due to atmospheric influence, resulting in inflated seasonal patterns. Additionally, it shows low concentrations during summer, which is not correct if compared to field data. The L2G SR product (black line) provides intermediate chlorophyll-a estimates characterized by more gradual seasonal variations, yet it continues to overstate peak concentrations in comparison to field data. The suggested methodology (red line) exhibits a more consistent and precise depiction of seasonal phenological patterns, closely aligning with established seasonal dynamics of chlorophyll-a in coastal and inland waters. The findings for the Ambracian Gulf underscore seasonal patterns, featuring a slow rise in spring, presenting the phenological bloom. The highest chlorophyll-a concentrations are presented in August, reaching up to 5 mg/m³ and revealing hypertrophic trophic conditions. This seasonal study emphasizes the ability of the proposed technique to mitigate sensor-induced biases and accurately capture ecologically significant chlorophyll-a dynamics. The suggested method addresses the shortcomings of conventional satellite products, providing a reliable tool for the long-term assessment of water quality and phenological trends.

3.4. Replicate Analysis in the Aitoliko Lagoon

Figure 7a illustrates the temporal variability of chlorophyll-a concentration (mg/m³) over the period 2002–2023, highlighting the differences between the two datasets. The suggested approach reliably produces smoother trends while preserving essential peaks (Figure 7c). Figure 7b depicts the interannual trends of chlorophyll-a concentration, averaged annually (regression lines: L3: $1.653-0.00223(year-2002)$, p-value for the slope: 0.779, L2G: $4.38+0.06057(year-2002)$, p-value for the slope: 0.005, proposed: $3.38+0.03351(year-2002)$, p-value for the slope: <0.001), showing that the suggested technique reduces variability without considerable loss of peak values relative to the L2G dataset. In addition, the proposed methodology successfully identifies a positive trend that aligns with the initial L2G dataset. After the pre-processing procedure, 25.65% of the dataset was attributed as “NA”. Phenological trends were successfully identified in the proposed methodology but might not be clear in the initial dataset due to errors produced by cloud coverage and other observational discrepancies. The algal bloom in Aitoliko occurs from May to September, peaking in August, at up to 2.5 mg/m³. This outcome highlights the importance of the study in ensuring the identification of phenological patterns and enabling evaluations of climate change effects on aquatic ecosystems. This study confirms previous findings from the Ambracian Gulf, demonstrating the efficacy of the suggested methodology in the replicable study area.

4. Discussion

The results of this study demonstrate the effectiveness of the suggested methodology in overcoming chlorophyll-a phenological errors using L2G surface reflectance products. The suggested dataset exhibits improved preservation of phenological patterns and reduced noise from factors such as cloud covering through the implementation of statistical corrections, resolution of logical errors, and use of a LOESS filter. These improvements have proven critical for monitoring seasonal and interannual variations in chlorophyll-a concentrations, which are crucial markers of ecological health and climate-induced alterations in aquatic environments [48]. Chlorophyll-a concentration varies throughout the year due to seasonal changes in temperature, light availability, and nutrient levels, which influence phytoplankton growth. During spring and summer, increased sunlight and warmer temperatures promote photosynthesis, leading to higher chlorophyll-a concentrations, especially in nutrient-rich waters, where upwelling or runoff provides essential nutrients like nitrogen and phosphorus. In contrast, during autumn and winter, reduced sunlight, lower temperatures, and increased water mixing can limit phytoplankton growth, causing chlorophyll-a levels to decline.

The LOESS algorithm has previously been used in remote sensing studies with different applications, including reconstructing missing storage data [55], modeling satellite data [56], calibrating the satellite alignment angle [57], mitigating noise mitigation in Sentinel-1 data [58], and so forth. Compared to other smoothing algorithms, the LOESS application is a non-parametric algorithm [57] that can effectively manage outliers, since it gives more weight to points closer to the target of estimation [43], which is beneficial in environmental data that may contain noise. Other smoothing models such as Simple Moving Average [59] are limited to linear patterns and can mask variability, making them unable to detect the seasonal patterns of chlorophyll-a. In addition, comparing the LOESS application to polynomial regression, which both can model data trends, but the polynomial regression may not adequately capture local variability [60]. Finally, compared to seasonal decomposition using classical methods, the LOESS offers more flexibility and better performance in handling varying seasonal patterns, since it allows the seasonal component to adapt over time [61].

Furthermore, in this study, the proposed methodology using the LOESS algorithm successfully addresses the identification of seasonal patterns. The time series results demonstrate that there might be an underestimation in the derived chlorophyll-a concentration, but the seasonal peak was successfully identified. While the method copes well with the high percentage of missing values (46% of the chlorophyll-a dataset), further investigation should be conducted for different percentages of missing values. The study aims to perform LOESS smoothing while identifying chlorophyll-a’s phenological patterns and seasonal distribution. The LOESS algorithm presents true phenological patterns, relies on a large dataset, and is sensitive to outliers. Moreover, the successful implementation of the LOESS algorithm is based on the exclusion of extreme outliers with no physical meaning (deriving from sensor errors and cloud coverage) during the pre-processing procedure. Additionally, the smoothing span parameter was selected to ensure the identification of seasonal phenological patterns and, at the same time, local variations were still tracked relatively closely. In summary, the limitation of the LOESS algorithm is its sensitivity to abrupt changes [57], a phenomenon that is eliminated by daily temporal resolution of the dataset and the implementation of pre-processing logical corrections. In addition, it has been discussed that the LOESS algorithm can struggle with producing accurate predictions at the endpoints of the data range, known as edge effects [56]. This difficulty has been eliminated due to the long time series dataset, which increases computational time [40] but provides high accuracy in the produced dataset.

The study confirms the applicability of this methodology to various aquatic systems, as evidenced by the replication of previous findings from the Ambracian Gulf in the Aitoliko Lagoon. The study reveals that the phenological bloom of the Ambracian Gulf occurs during the summer period. Field investigations from 2009 to 2010 [32] revealed spring–summer phytoplankton blooms, indicating mesotrophic to eutrophic conditions. Finally, the study is in accordance with the results of the Water Framework Directive, which has classified the Ambracian Gulf (station GR000500010002N) as having a “bad” trophic state since the trophic range is higher than 2.21 mg/m³ [62].

In addition, the Aitoliko Lagoon exhibits variable trophic conditions, transitioning from mesotrophic to hypertrophic states (Figure 6d and Figure 7b,c). Between May 2006 and April 2007, chlorophyll-a concentrations ranged from 3.26 to 14.89 mg/m³, signifying mesotrophic conditions [63]. Similar to the results of our analysis, eutrophic to hypertrophic conditions have been seen, especially in the summer months (July–August), characterized by cyanobacterial blooms linked to increased nutrient concentrations. Comparable findings were documented in further investigations done from July 2013 to May 2014, and more recently from February 2023 to August 2023, corroborating the lagoon’s mesotrophic state under certain conditions [11]. Chlorophyll-a concentrations ranging from 2.5 to 7 mg/m³ were observed in May 2022, reinforcing mesotrophic conditions [12]. Finally, phytoplankton blooms from June to August in 2013–2014 further demonstrated the lagoon’s eutrophic tendencies, along with water column stratification and sedimentological alterations [64]. These findings highlight the interaction between natural processes and human-induced stresses in influencing the ecological dynamics of water bodies.

Moreover, the research demonstrates that although L3 datasets can yield interesting insights, they frequently exhibit considerable uncertainty, particularly in inland and coastal waters [65]. L3 datasets may be undermined by geolocation inaccuracies and temporal inconsistencies, resulting in notable disparities. Comparison with field data measurements presented here underlines the problems generated by the text when confronted with field data and L2 datasets. Researchers agree that L3 chlorophyll-a datasets can provide significant information, but their trustworthiness depends upon the methods utilized and their suitability for the specific maritime environment [66,67]. Emphasis is given to the necessity of local calibration and adjustment of these algorithms to enhance precision and minimize errors in chlorophyll-a retrievals.

Therefore, field data measurements from the Ambracian Gulf provide decisive criteria to calibrate algorithms for remote sensing analysis and highlight the efficacy of the methodology in mitigating sensor-related biases, especially in intricate aquatic systems where cloud coverage and water turbidity pose considerable challenges. The suggested methodology offers a robust framework for monitoring water quality over extended durations by calibrating the methodology using field data and verifying it across various temporal scales. Researchers have identified the need for field data measurements and suggested different methods to correct remote sensing data and enhance data quality [11,68,69]. Remote sensing datasets can provide valuable feedback for water bodies’ quality status monitoring in near-real time, because field data measurements are very difficult and very costly. Implementing the proposed methodology in L2 datasets as described earlier, we can sufficiently overcome the shortcomings of the L2 datasets and successfully assess the phenological patterns and trends of water bodies.

5. Conclusions

Remote sensing technologies address global challenges in a creative, cost-efficient, timesaving, and environmentally sustainable way. In recent years, we have witnessed significant and swift gains in remote sensing analysis inspired by enhancements in data accessibility, computational methods, and algorithmic innovations. The suggested methodology aims to enhance the produced chlorophyll-a datasets by providing straightforward statistical analysis combined with a conventional chlorophyll-a retrieval method. The methodology studied effectively addresses logical errors, applies statistical corrections, and implements the LOESS criteria to fill in the missing values caused by cloud interference while preserving phenological patterns, thus enhancing the accuracy and reliability of remote sensing-based environmental monitoring.

Lastly, comparing LOESS with other correction methods could be an interesting aspect of future research and help showcase the actual role of LOESS in handling analogous problems. Furthermore, water bodies’ ecosystem dynamics could also be considered in a comprehensive uncertainty analysis to examine how the proposed methodology affects the uncertainty of chlorophyll-a concentration.