EVALUACIÓN DE ÍNDICES BIO-ÓPTICOS HIPERESPECTRALES PARA LA CLASIFICACIÓN DE CLOROFILA-A EN AGUAS COSTERAS – BAHÍA DE SECHURA, PERÚ

1. INTRODUCTION 🡩

Estimating chlorophyll-a (Chl-a) concentrations in optically complex coastal waters remains a significant challenge in ocean color remote sensing due to the combined influence of colored dissolved organic matter (CDOM), total suspended matter (TSM), and non-algal particles on the water’s output reflectance signal. Traditional empirical algorithms, based on the relationships between blue and green bands, have been widely applied in clear open ocean waters, but their performance drops considerably in coastal environments, where CDOM and inorganic sediments increase spectral interference in the blue domain (Mishra & Mishra, 2012).

To address these limitations, red and near-infrared (NIR) spectral approaches have been developed, exploiting the phytoplankton absorption feature near 665 nm and the reflectance peak around 700 nm. Among the most widely used methods, the Normalized Difference Chlorophyll Index (NDCI) proposed by Mishra and Mishra (2012) showed strong performance in productive and turbid waters (R² = 0.95 in simulations and R² = 0,90 when applied to MERIS imagery from diverse coastal systems, including Chesapeake Bay and the Mississippi River Delta). Similarly, the three-band model of Gitelson et al. (2009) reduces TSM and CDOM interference using reflectance at 665 nm and an NIR reference band, achieving R² values above 0,94 in eutrophic inland and coastal waters.

The diversity of optical water types observed in coastal environments has motivated the development of classification-based approaches that assign chlorophyll-a (Chl-a) estimation algorithms according to the dominant optical regime, rather than applying a single universal model. Tran et al. (2023) evaluated band-ratio inversion models for estimating chlorophyll-a, from Sentinel-2 MSI (MultiSpectral Instrument) and Sentinel-3 OLCI (Ocean and Land Colour Instrument) imagery, using an extensive in situ dataset of reflectance and chlorophyll-a measurements. They classified five Optical Water Types (OWTs) and demonstrated that blue/green models are suitable for clear to moderately turbid waters, whereas red/NIR models perform better in turbid environments with high Chl-a concentrations.

Despite substantial advances in hyperspectral methods for chlorophyll-a retrieval and classification, studies addressing the bio-optical characteristics of Peruvian coastal waters remain limited. In particular, the use of in situ hyperspectral observations to classify Chl-a concentrations into ecologically relevant categories has received little attention, despite its potential value for monitoring ecosystem productivity, characterizing phytoplankton dynamics, and detecting anomalous environmental conditions.

To help address this knowledge gap, three oceanographic field campaigns were conducted in Sechura Bay, Piura, between 2024 and 2025 as part of a collaborative research program between the Instituto del Mar del Perú (Imarpe) and the Korea Institute of Ocean Science and Technology (KIOST). These surveys collected hyperspectral radiometric measurements and complementary environmental data to characterize the surface bio-optical properties of this highly productive coastal ecosystem.

The present study aims to classify chlorophyll-a concentrations using bio-optical indices derived from in situ hyperspectral data. Specifically, we investigate the optical contribution and discriminatory capability of individual indices with respect to the Chl-a spectral signal and subsequently evaluate their performance using supervised classification techniques.

2. MATERIALS AND METHODS 🡩

Study area

Sechura Bay is located on the northern coast of Peru (5°05′–5°20′ S, 80°51′–81°10′ W) within the Piura region and encompasses approximately 89 km of coastline (Fig. 1). The bay supports extensive artisanal fisheries targeting benthic and demersal resources, as well as intensive aquaculture activities. Its ecological dynamics are seasonally influenced by adjacent coastal wetlands, including the Virrilá Estuary, the San Pedro and Palo Parado mangrove systems, and the Piura River, whose influence becomes particularly pronounced during El Niño events.

The hydrography of Sechura Bay is shaped by its open-ocean exposure and bathymetric configuration. Throughout most of the year, the bay is influenced by the interaction between warm northern waters and the colder waters associated with the Peruvian Coastal Current. These conditions promote a circulation pattern characterized by the intrusion of subsurface waters (<10 m depth) through the central sector of the bay, followed by southeastward transport toward Matacaballo and the Virrilá Estuary, with current velocities reaching up to 30 cm/s (Morón et al., 2013).

Field campaigns were conducted on 23 -24 April 2024, 14 -15 August 2024, and 20 -22 May 2025. Sampling stations extended from nearshore waters to approximately 8 nautical miles offshore. Each survey included 13 stations occupied aboard a wooden artisanal vessel. At each station, hyperspectral radiometric measurements were acquired and water samples were collected for chlorophyll-a (Chl-a), phytoplankton abundance, and total suspended solids (TSS) analyses. In addition, water transparency, sea surface temperature (SST), and sea surface salinity (SSS) were measured (Table 1).

Radiometric Measurements

Above-water radiometric observations were collected using three TriOS Ramses hyperspectral radiometers covering the spectral range from 350 to 900 nm. Measurements were recorded every 5 s over a 5-min sampling period, resulting in 2,332 valid reflectance spectra obtained from 38 sampling stations. The radiometers were mounted on a steel frame positioned alongside the vessel to minimize platform-induced shading effects (Espinoza Villar et al., 2013). Instrument deployment followed the acquisition geometry recommended by Mobley (1999). The RAMSES-ACC sensor, equipped with a cosine collector for downward-directed light detection (α ≤ 90°), was mounted atop the structure facing the zenith. This sensor measured total incident irradiance above the water surface (E_s(λ)), expressed in W m^-2. Two additional RAMSES-ARC sensors, each with a 7° field of view (FOV), were used to measure water-leaving radiance (L_t (λ, θ, ϕ)) and sky radiance (L_s (λ, θ, ϕ)) in W m^-2 sr^-1 (Ade et al., 2023; Fig. 2). With the Sun as reference, the nadir (θ) and azimuth (ϕ) angles were set to 40° and 135°, respectively, for L_t, and to 140° and 135°, respectively, for L_s (da Silva et al., 2021). Remote sensing reflectance (R_rs) was subsequently calculated using the following equation:

Where ρ is a factor related to the fraction of sky radiance (L_s) reflected at the air–water interface toward the sensor, and is a function of wind speed, solar zenith angle, and the sensor's zenith and azimuth angles. A reflection factor of 0.028 was applied, as this value is appropriate for the acquisition geometry used and for wind speeds below 5 m/s under clear blue skies, or for any wind speed under overcast conditions (Mobley, 1999; da Silva et al., 2021).

chlorophyll-a Determination

chlorophyll-a concentrations were determined for samples collected at the 38 monitoring stations following the fluorometric protocol described by Knap et al. (1996) using a Turner Designs 10-AU fluorometer. Briefly, 200 mL seawater samples were filtered through Whatman GF/F glass-fiber filters (0.75 μm nominal pore size) to retain phytoplankton pigments. Filters were stored in glassine paper envelopes or aluminum foil, labeled, and subsequently homogenized in 10 mL of 90 % acetone using either a grinding rod or ultrasonic disruption. Samples were extracted for 24–72 h and then centrifuged at 4,500 rpm for 5 min. The resulting supernatant was analyzed fluorometrically. A first measurement was obtained before acidification (Fo) followed by a second measurement after adding 0.1 mL HCl or two drops of 10 % HCl solution (Fa). Prior to each measurement, fluorometer cuvettes were rinsed with at least 2 mL of 90 % acetone. The Fo and Fa readings were subsequently used to calculate chlorophyll-a concentration.

Sea Surface Temperature and Salinity

Sea temperature and salinity profiles were measured using a CastAway-CTD profiler. Measurements were acquired at the surface and throughout the water column according to the depth of each sampling station.

Method

The spectral classification workflow began with quality control of the hyperspectral reflectance data. Spectra affected by vessel motion or exhibiting negative R_rs<0 values within the selected spectral range (400–750 nm) were discarded. A representative reflectance spectrum was then selected for each station based on the coefficient of variation (CV%) computed at each wavelength (λ), allowing the appropriate use of either the mean or median spectrum according to the variability observed within each dataset (Fig. 3).

Subsequently, several bio-optical indices (Table 2) were calculated from R_rs (λ), to characterize chlorophyll-a and other optically active constituents. These included the Normalized Difference Chlorophyll Index (NDCI), band ratios BR_705/665, BR_665/490, the Fluorescence Line Height (FLH), the Green Index (GI), and a three-band algorithm. Relationships between Chl-a concentration and bio-optical indices were assessed using Spearman rank correlation analysis.

The resulting indices served as predictor variables for three supervised classification models: Linear Discriminant Analysis (LDA; Fisher, 1936), Random Forest (RF; Breiman, 2001), and Partial Least Squares Discriminant Analysis (PLS-DA; Wold et al., 2001). The six bio-optical indices derived from R_rs (λ) constituted the predictor set for all models. Model performance was evaluated using leave-one-out cross-validation (LOOCV). Classification accuracy, Cohen’s Kappa coefficient (Landis & Koch, 1977), sensitivity, and specificity were calculated for each chlorophyll-a class. Satellite-image processing and visualization were conducted using SEADAS, SNAP, and QGIS, whereas statistical analyses were performed in RStudio.

3. RESULTS 🡩

3.1 Surface Reflectance Signatures

The hyperspectral reflectance measurements collected during the three field surveys conducted in Sechura Bay between 2024 and 2025 revealed pronounced spatial and temporal variability in surface optical properties. Variations in spectral shape and reflectance magnitude were associated with changes in chlorophyll-a concentration, phytoplankton abundance, total suspended solids, and environmental conditions such as incident solar radiation and cloud cover.

During the April 2024 survey, stations located in the coastal sector of the bay (stations 1, 4, 5, 8, and 13) exhibited a characteristic bimodal reflectance pattern, with a primary maximum between 560 and 590 nm and a secondary peak between 670 and 690 nm. This spectral configuration is commonly associated with productive waters characterized by elevated phytoplankton biomass and increased chlorophyll-a concentrations. In contrast, stations situated in the central and outer sectors of the bay (stations 2, 7, 11, and 12) displayed lower reflectance values and weaker spectral features, indicative of reduced biological productivity (Fig. 4).

The August 2024 survey exhibited a markedly different optical regime. Most stations showed relatively weak reflectance signals, consistent with lower phytoplankton biomass and chlorophyll-a concentrations. Only stations 1 and 13 retained spectral characteristics suggestive of enhanced productivity. In particular, the absence of a distinct reflectance maximum between 670 and 700 nm at stations 3, 6, 7, 8, and 9 suggests limited phytoplankton development and reduced Chl-a content (Fig. 5).

The optical characteristics observed during August coincided with elevated concentrations of suspended particulate matter. TSS values ranged from 16.18 to 27.86 mg/L at seven coastal stations, exceeding those measured during April, when the maximum concentration reached 23.90 mg/L. The broad low-reflectance plateau extending from approximately 450 to 580 nm is consistent with enhanced light scattering by suspended particles. The spatial distribution of TSS suggests that local circulation patterns play a key role in controlling sediment transport and retention within the bay.

The highest levels of biological productivity were observed during May 2025. Coastal stations 1, 8, 9, 11, and 13 exhibited well-defined reflectance maxima near 580 and 690 nm. The increase in reflectance around 580 nm, with values exceeding 0.0035 sr^-1, indicates elevated phytoplankton biomass. However, the relatively modest enhancement near 690 nm suggests moderate chlorophyll-a concentrations at several stations, including stations 3, 5, 7, and 10 (Fig. 6).

Elevated suspended matter concentrations were also observed during May, particularly in the northern sector of the bay. Stations 7, 9, 10, 11, and 13 recorded mean TSS concentrations of approximately 31.10 mg/L. These observations were consistent with reduced water transparency and increased attenuation of blue wavelengths, particularly around 480 nm.

The analysis of hyperspectral signatures allowed the identification of distinct bio-optical regions within Sechura Bay. A highly productive coastal zone was identified around stations 1, 4, 8, 9, and 13, where elevated chlorophyll-a concentrations coincided with spectral features characteristic of phytoplankton-rich waters. The productivity of this sector is likely enhanced by seasonal freshwater and nutrient inputs from the Piura River and the Virrilá Estuary. An intermediate-productivity region was identified in the central portion of the bay around stations 10, 11, and 12, whereas stations 2, 3, 5, 6, and 7 consistently exhibited spectral characteristics associated with comparatively low biological productivity.

3.2 Oceanographic Conditions in Sechura Bay

Oceanographic conditions varied seasonally throughout the study period. Mean sea surface temperature (SST) was 18.48°C during April 2024 and decreased to 17.17°C in August 2024, reflecting the influence of cooler winter conditions. By May 2025, SST increased to 19.38°C. In contrast, sea surface salinity (SSS) remained relatively stable, ranging from 34.85 to 35.11.

Marked temporal differences were observed in suspended particulate matter concentrations. The highest TSS levels occurred during May 2025, when eight stations exceeded 18 mg/L. Five of these stations were located in the northern sector of the bay, where concentrations averaged approximately 29 mg/L, likely reflecting the influence of freshwater discharge and sediment transport from the Piura River. During August 2024, the highest TSS concentrations were recorded in the southern sector of the bay, averaging approximately 24 24 mg/L. In contrast, April exhibited comparatively lower TSS concentrations, ranging from 11.88 to 23.90 mg/L.

3.3 chlorophyll-a Classification

chlorophyll-a concentrations ranged from 0.36 to 13.84 mg/m³, with the highest values generally occurring in coastal waters. These concentrations are consistent with oligotrophic to mesotrophic conditions commonly reported for Sechura Bay. During April 2024 and May 2025, Chl-a concentrations ranged from 2.2 to 13.84 mg/m³, whereas during August 2024 concentrations remained below 1 mg/m³ at most stations. The highest concentration was recorded at station 13 (4,42 mg/m³) (Fig. 7).

The spatial distribution of chlorophyll-a was strongly reflected in the hyperspectral measurements. Stations with Chl-a concentrations be-low 3 mg/m³ generally exhibited remote-sensing reflectance values lower than 0.003 sr^-1, whereas stations with elevated chlorophyll-a concentrations showed reflectance values exceeding 0.006 sr^-1 near 690 nm. Similar trends were observed in phytoplankton abundance, reinforcing the biological significance of the spectral response.

During April, most spectra exhibited two characteristic reflectance maxima: a broad peak between 550 and 600 nm and a second maximum near 700 nm. These spectral features are typical of phytoplankton-dominated waters and closely resemble the signatures described by Schalles (2006) for several phytoplankton taxa. Among them, the spectral pattern attributed to Navicula sp. showed the greatest similarity to the reflectance signatures observed in Sechura Bay, consistent with its occurrence across all sampling stations.

For classification purposes, in situ Chl-a concentrations were divided into three categories --low, moderate, and high --using the 33rd and 67th percentiles of the observed distribution (p₃₃ = 2.35 mg/m³; p₆₇ = 7.36 mg/m³). This approach produced three nearly balanced groups comprising 13, 12, and 13 observations, respectively. Mean chlorophyll-a concentrations differed markedly among classes, averaging 1.036 mg/m³ for the low class, 4.337 mg/m³ for the moderate class, and 9.972 mg/m³ for the high class. The clear separation among class means supports the use of these thresholds for supervised classification (Table 3). Moreover, in investigations with limited data, a minimum of 10 observations per class is required to estimate model errors when evaluating reliability and accuracy. The coefficient of variation (CV) varies considerably among classes: CV_low=51 %, CV_moderate=29 %, and CV_higha=24 %. The elevated CV observed in the low class is characteristic of very low Chl-a concentrations, where minor absolute differences translate into high relative variability. In contrast, the high class exhibits the greatest relative homogeneity, which enhances its spectral detectability.

For all three classes, median values were lower than the corresponding means, indicating positively skewed distributions. This statistical characteristic supports the use of the median rather than the mean as the representative spectral signature for each station, differing from the standard NASA Ocean Color protocol commonly applied to Case II coastal waters (Mueller et al., 2003). Furthermore, the class thresholds derived from the percentile approach correspond closely to transitions between mesotrophic and eutrophic trophic states reported for tropical coastal ecosystems (Vollenweider & Kerekes, 1982; Carlson, 1977), providing ecological justification for the classification framework. The resulting balanced class structure also ensured adequate sample representation for leave-one-out cross-validation (LOOCV), thereby enhancing the robustness of model performance assessment.

3.4 Performance of Bio-Optical Indices

All six bio-optical indices successfully discriminated among the three Chl-a concentration classes, indicating that the classification scheme captured meaningful differences in the optical properties of Sechura Bay waters (Fig. 8). The Kruskal–Wallis H statistic calculated for the Green Index (GI) shows a systematic pattern: maximum values at the 13 stations falling into the high Chl‑a class, minimum values at the 13 low‑class stations, and intermediate values at the 12 medium‑class stations (H = 16.624; Table 4).

The superior performance of GI was corroborated by the correlation analysis, which revealed the strongest association with in situ Chl-a concentrations (ρ = 0.701, p < 0.001). This result exceeded those obtained for the NDCI and BR_705/665 indices, suggesting that, within Sechura Bay, phytoplankton reflectance manifests mainly in the green band (555 nm) of the visible spectrum.

The NDCI, BR_705/665, and 3_band indices exhibited identical discriminatory performance (H = 14.784, adjusted p = 0.00092), reflecting their shared dependence on the R_{rs 665} and R_rs 705 bands. Although BR_665/490 and FLH_proxy also showed statistically significant separation among classes, their discriminatory power was comparatively weaker (H = 13.151 y 10.530, respectively; adjusted p < 0.01). In particular, the reduced performance of the FLH_proxy probably results from scattering induced attenuation of the fluorescence signal in optically complex coastal waters when broad spectral bands are used for estimation. These findings demonstrate that in situ R_rs spectra acquired with the TriOS Ramses radiometer provide statistically reliable spectral information for distinguishing Chl-a concentration classes in Sechura Bay. Incorporating measurements of co varying optically active constituents—such as CDOM and suspended inorganic particulate matter—in future studies may further enhance classification accuracy.

A Spearman correlation analysis showed that all six bio-optical indices were positively associated with Chl-a concentration, with correlation coefficients ranging from 0.607 to 0.701 (ρ = 0.607-0.701, p < 0.001; Table 4). The Green Index (GI) yielded the highest correlation (ρ = 0.701), exceeding the performance of red/NIR based indices that are conventionally recommended for turbid coastal waters.

The NDCI, el BR₇₀₅/₆₆₅, and three band index yielded identical correlation coefficients (ρ = 0.645), which aligns with their common mathematical reliance on the 665 nm and 705 nm bands. The FLH index, despite being theoretically specific to active chlorophyll-a fluorescence, produced the lowest correlation (ρ = 0.607), probably as a result of signal attenuation in optically complex coastal waters. The GI captures the displacement of the reflectance peak toward the green band with increasing Chl-a concentration—a typical feature of Case II waters containing mixtures of phytoplankton, CDOM, and suspended sediments, as documented in recent studies. The consistently positive correlations (an entirely red correlation matrix) demonstrate that all indices respond to the same bio optical signal, supporting the use of GI as a parsimonious model representative (Fig. 9).

The Spearman correlation coefficients obtained (ranging from 0.607 to 0.701) theoretically predict an accuracy of 60–65 % (Table 4), which matches the observed model performance (RF and PLS = 60.5 %, LDA = 63.2 %). This indicates that the models are extracting virtually all the information embedded in the indices, with no evidence of significant underfitting. Classification performance is constrained by class overlap, most notably in the medium class. Additionally, Figure 10a illustrates the relationship between the GI index and Chl‑a concentration, which is well described by a linear regression (R²=0.406, ρ=0.701, p<0.0001).

Mishra and Mishra (2012) noted that two band indices employing red and NIR wavelengths are strongly correlated with chlorophyll a. Nevertheless, the substantial influence of SST and CDOM in more turbid waters restricts the applicability of such indices to relatively clear environments. Sechura Bay receives sediment inputs throughout most of the year, creating moderately turbid conditions that favor the green index over red/NIR alternatives. This outcome aligns with the blue/green ratio algorithms OC3, OC5, and OC6 reported by Tran et al. (2023), who achieved R² values of 0.57 and 0.61 for OWT types 1 and 3, respectively.

Cesar et al. (2023) showed that ocean color algorithms for Chl-a estimation performed slightly better when using the three band OC3 algorithm, specifically by selecting the band ratio least affected by CDM (CDOM + detritus) absorption (R_rs 490/555). Conversely, the GI₅₅₅/₄₉₀ represents the mathematical inverse of the OC2 algorithm, which NASA has used operationally for SeaWiFS (O'Reilly & Werdell, 2019) and which yields the highest R² among two‑band indices globally. This supports the optimal use of the green index in Sechura Bay, as both indices capture the same blue‑green spectral contrast with comparable signal strength—consistent with the pattern shown in Figure 10b.

3.5 Classification Models

Among the three supervised classification approaches evaluated, LDA achieved the highest overall performance, with an accuracy of 63.2% and a Cohen's Kappa coefficient of 0.446. According to the classification scheme proposed by Landis and Koch (1977), this level of agreement falls within the “moderate” range (0.41 -0.60), which is generally considered acceptable for coastal remote-sensing applications. This outcome is expected for correlated variables, since the number of parameters to be estimated is small, resulting in lower estimation variance. LDA attains the highest overall accuracy (63.2%) and demonstrates the most balanced class_wise performance, with sensitivity values of 84.6% for the low class, 41.7% for the moderate class, and 61.5% for the high class. These results justify selecting LDA as the recommended model. In contrast, Random Forest (RF) and PLS_DA yield identical overall accuracies (60.5%) but exhibit distinct error distributions across classes (Table 5).

Sensitivity analyses showed that all models classified the low-Chl-a class with high reliability, followed by the high-Chl-a class. In contrast, the moderate-Chl-a category proved considerably more difficult to identify. This class encompassed concentrations ranging from 2.72 to 7.23 mg/m³, overlapping substantially with the percentile thresholds used to define class boundaries (p₃₃ = 2.35, p₆₇ = 7.36 mg/m³). Consequently, it represents a spectral transition zone characterized by considerable ambiguity. Furthermore, its relatively high coefficient of variation (CV = 28.5 %) indicates greater internal heterogeneity, as stations with Chl-a concentrations between approximately 3 and 7 mg/m³ did not exhibit spectrally consistent signatures (Fig. 11).

All models exhibited high specificity, ranging from 69 % to 92 %, particularly for the low-Chl-a class (84–92 %). This finding indicates a low probability of incorrectly assigning highly productive waters to the low-productivity category, an important characteristic for environmental monitoring programs focused on detecting ecological anomalies and algal bloom events.

3.6 Model Evaluation

The LOOCV results confirmed the robustness of the classification framework. LDA achieved the highest overall accuracy (63.2 %; κ = 0.446), followed by Random Forest and PLS-DA, both of which reached 60.5 % accuracy with Kappa values close to 0.40. Importantly, all models significantly outperformed the no-information rate of 34.2 % (p < 0.01), demonstrating that the classification skill was substantially greater than random assignment. A consistent pattern emerged across all models: the moderate chlorophyll class exhibited the lowest sensitivity (25.0–41.7 %). This result reflects the spectral overlap inherent to transitional chlorophyll concentrations (2,72 - 7,23 mg/m³), where the optical influence of phytoplankton is neither sufficiently weak nor sufficiently strong to dominate the reflectance signal. In this range, the contribution of chlorophyll competes with absorption and scattering processes associated with CDOM and suspended particulate matter, both of which are characteristic features of Case II coastal waters.

PLS-DA achieved the highest sensitivity for the low-Chl-a class (92.3 %), although this improvement occurred at the expense of markedly poorer performance for the intermediate class. In contrast, LDA maintained a more balanced distribution of classification errors across all categories.

These findings closely resemble those reported by Watanabe et al. (2020), who applied machine-learning algorithms (ANN, RF, and SVM), with R_rs (λ) data, to classify trophic conditions in Brazilian reservoirs using hyperspectral reflectance and in situ chlorophyll observations. Their results showed that extreme productivity states are generally easier to classify than intermediate trophic conditions. A comparable pattern was observed in Sechura Bay, where the low- and high-Chl-a classes exhibited substantially higher classification success than the transitional intermediate class.

The spectral basis for this behavior is illustrated in Fig. 12. The three classes are associated with clearly differentiated spectral signatures in the visible R_rs (λ) domain. Thin lines correspond to the mean spectra of individual stations, while thick lines represent the class median. Vertical dashed lines mark key bio-optical wavelengths (443, 490, 555, 665, and 705 nm). Class boundaries are based on the 33rd and 67th percentiles: low class < 2.35 mg/m³ (blue), moderate class between 2.35 and 7.36 mg/m³ (orange), and high class > 7.36 mg/m³ (red).

The low class is dominated by absorption from water and CDOM, displaying a weak chlorophyll a signal and peak reflectance in the blue green band. The moderate class acts as a transitional category, characterized by increasing reflectance in the red portion of the spectrum. In the high class, reflectance is driven by elevated algal biomass; stations with Chl-a concentrations exceeding 10 mg/m³ show a marked peak at 555 nm, where R_rs (λ) attains values as high as 1.5×10⁻².

4. DISCUSSION 🡩

The moderate classification accuracies obtained in this study (60–63 %) are comparable to those reported for optically complex coastal waters using in situ hyperspectral radiometry and datasets of similar size (Shi et al., 2013; Binding et al., 2013). The consistent difficulty across all four models in correctly classifying the medium chlorophyll a class indicates that this concentration range corresponds to an optical transition zone within the bay, where the observed spectral signature cannot be unequivocally assigned to phytoplankton biomass alone.

The performance of the LDA model (κ = 0.446) is comparable to results obtained in other Case II coastal waters, where the optical influence of CDOM and suspended inorganic particles reduces the sensitivity of chlorophyll-specific spectral indices. The consistently poor classification of the moderate Chl-a class, regardless of algorithm complexity or predictor selection, indicates that the limitation is primarily driven by intrinsic optical ambiguity rather than methodological shortcomings. This finding suggests that intermediate productivity conditions are characterized by overlapping spectral responses arising from the combined influence of phytoplankton, CDOM, and suspended sediments.

5. CONCLUSIONS 🡩

Among the bio-optical indices evaluated, the Green Index (GI) exhibited the strongest relationship with chlorophyll-a concentration, outperforming the red–near-infrared indices traditionally recommended for turbid coastal waters. This result indicates that the blue–green spectral contrast constitutes the dominant bio-optical signal in Sechura Bay. The enhanced performance of GI is likely associated with the optical influence of terrigenous inputs and colored dissolved organic matter, which attenuate the red-edge reflectance signal near 705 nm that is commonly observed in productive coastal ecosystems of the Humboldt Current System.

Linear Discriminant Analysis (LDA), using six bio-optical indices derived from hyperspectral remote-sensing reflectance, achieved the highest classification accuracy under leave-one-out cross-validation (LOOCV), outperforming both Random Forest and PLS-DA models despite their greater complexity. The superior performance of LDA demonstrates that linear relationships between bio-optical indices and chlorophyll-a classes are sufficient to capture the dominant patterns of optical variability within the available dataset. These findings further indicate that increasing model complexity does not necessarily improve classification performance when sample size is limited and predictor variables are highly correlated.

A consistent outcome across all evaluated models was the reduced sensitivity associated with the moderate Chl-a class. Because this pattern persisted regardless of the classification algorithm or predictor set employed, the difficulty in identifying intermediate concentrations should not be interpreted as a methodological limitation. Instead, it reflects an intrinsic optical characteristic of transitional waters, where the simultaneous influence of phytoplankton biomass, CDOM, and suspended particulate matter generates overlapping spectral signatures that cannot be fully resolved using reflectance-based indices alone.

The hyperspectral remote-sensing reflectance measurements, R_rs (λ), acquired with the TriOS Ramses radiometers successfully discriminated chlorophyll-a concentration classes in the optically complex coastal waters of Sechura Bay. These results establish a quantitative bio-optical framework for future satellite-validation activities and provide a methodological foundation for integrating field radiometry and satellite observations into long-term phytoplankton biomass monitoring programs along the Peruvian coast.

O’Reilly, J. E., Maritorena, S., Mitchell, B. G., Siegel, D. A., Carder, K. L., Garver, S. A., Kharu, M. & McClain, C. (1998). Ocean color chlorophyll algorithms for SeaWIFS. Journal of Geophysical Research, 103(C11), 24937-24953. https://doi.org/10.1029/98JC02160

O’Reilly, J. E. & Werdell, P. J. (2019). Chlorophyll algorithms for ocean color sensors - OC4, OC5 & OC6. Remote Sensing of Environment, 229, 32-47. https://doi.org/10.1016/j.rse.2019.04.021

Schalles, J. F. (2006). Optical remote sensing techniques to estimate phytoplankton chlorophyll-a concentrations in coastal waters with varying suspended matter and CDOM concentrations. En L. L. Richardson & E. F. LeDrew (Eds.), Remote sensing of aquatic coastal ecosystem processes. Remote sensing and digital image processing (Cap. 3, pp.27-79). https://doi.org/10.1007/1-4020-3968-9_3

Shi, K., Li, Y., Li, L., Lu, H., Song, K., Liu, Z., Xu, Y. & Li, Z. (2013). Remote chlorophyll-a estimates for inland waters based on a cluster-based classification. Science of the Total Environment, 444, 1-15. https://doi.org/10.1016/j.scitotenv.2012.11.058

Tran, M. D., Vantrepotte, V., Loisel, H., Oliveira, E. N., Tran, K. T., Jorge, D., Mériaux, X. & Paranhos, R. (2023). Band ratios combination for estimating chlorophyll-a from Sentinel-2 and Sentinel-3 in coastal waters. Remote Sensing, 51, 1653. https://doi.org/10.3390/rs15061653

Vollenweider, R. A. (1992). Coastal marine eutrophication: principles and control. En R. A. Vollenweider, R. Marchetti & R. Viviani (Eds.), Marine coastal eutrophication. proceedings of an international conference (pp. 1–20). Elsevier Science. https://doi.org/10.1016/B978-0-444-89990-3.50011-0

Vollenweider, R. A. & Kerekes, J. (1982). Eutrophication of waters: Monitoring, assessment and control (Report of the Cooperative Programme on Eutrophication). Organization for Economic Co-Operation and Development (OECD). https://acortar.link/KnjpNe

Watanabe, F. S., Miyoshi, G. T., Rodrigues, T. W., Bernardo, N. M., Rotta, L. H., Alcântara, E. & Imai, N. N. (2020). Inland water’s trophic status classification based on machine learning and remote sensing data. Remote Sensing Applications: Society and Environment, 19, 100326. https://doi.org/10.1016/j.rsase.2020.100326

Wold, S., Sjöström, M. & Eriksson, L. (2001). PLSregression: a basic tool of chemometrics. Chemometrics and Intelligent Laboratory Systems, 58, 109–130. https://doi.org/10.1016/S0169-7439(01)00155-1

Agradecimientos

Un especial agradecimiento al co-director de KOPELAR-IMARPE, PhD. Han-Jun Woo por su apoyo en la financiación del proyecto piloto “Estudio de las propiedades bio-ópticas del agua y su relación con la actividad pesquera-acuícola en la bahía de Sechura a través de sensores remotos”. Asimismo, agradecer al ingeniero Manuel Guzmán Roca por el análisis de las muestras de sólidos suspendidos totales.

Rol de los autores

Carlos Paulino: conceptualización, investigación, redacción-borrador original, metodología, análisis formal, validación, administración del proyecto. German Velaochaga: redacción-borrador original, análisis formal, investigación. Marceliano Segura: metodología, redacción-revisión y edición, investigación. Joo-Hyung Ryu: investigación, metodología, redacción-revisión y edición, obtención de financiación. Jun-Ho Lee: conceptualización, investigación, redacción-revisión y edición, análisis formal y obtención de financiación. Jesús Ledesma: redacción-revisión y edición, investigación.

Declaración de conflicto de intereses

Los autores declaramos no tener conflictos de intereses financieros ni personales que puedan influir en los resultados de esta investigación.

Declaración de uso de Inteligencia Artificial

Durante la preparación de esta investigación, se utilizó Deep L para verificar la gramática y la ortografía de la redacción. Después de utilizar esta herramienta, como autores editamos, revisamos y aprobamos el resultado final; asumiendo la total responsabilidad por el contenido de la publicación.

Fuentes de financiamiento

Esta investigación recibió subvención del Korea Institute of Ocean Science and Technology (KIOST) para llevar a cabo tres operaciones en el mar como parte del proyecto piloto “Estudio de las propiedades bio-ópticas del agua y su relación con la actividad pesquera-acuícola en la bahía de Sechura a través de sensores remotos”, en el marco del convenio entre el Instituto del Mar del Perú (Imarpe) y el Korea Institute of Ocean Science and Technology (KIOST).