Field-Spectroradiometric Characterisation of Three Seagrass Species (Halophila stipulacea, Halodule uninervis, and Halophila ovalis) and Their Differentiation in the Arabian Gulf, Kingdom of Bahrain

Highlights

What are the main findings?

• Halophila ovalis is spectrally distinct from co-occurring Bahraini taxa across 692–1394 nm, with large effect sizes throughout the red-edge and near-infrared.
• A Sentinel-2-proxy linear discriminant analysis classifier reached 87.3% balanced accuracy, 20 percentage points above an equivalent Landsat-proxy model.

What are the implications of the main findings?

• Six species-specific Sentinel-2 vegetation indices, across two operational tiers, achieve balanced accuracies of 0.82–0.92.
• Classifier accuracy degrades by more than 7 percentage points at sensor noise σ ≈ 0.002 and by 5.5–15.7 percentage points at 2–3 nm wavelength shifts, requiring stringent calibration.

Abstract

Seagrass meadows support critical coastal ecosystems, but corresponding species-level remote sensing data remain limited, particularly in the Arabian Gulf, where field spectral data for dominant taxa are extremely limited. We present the first multi-species spectral characterisation of three dominant seagrass species in the Kingdom of Bahrain—Halophila stipulacea (n = 46 spectra, 25 stations), Halodule uninervis (n = 34, 19 stations), and Halophila ovalis (n = 17, 8 stations)—measured with an ASD FieldSpec^® 4 Hi-Res spectroradiometer (Malvern Panalytical, Malvern, UK; 350–2500 nm) from samples collected across 29 geographic stations (52 species–station sampling units). All sample counts reported here underwent quality control. Kruskal–Wallis tests with Benjamini–Hochberg (BH) correction, Jeffries–Matusita (JM) distance, Hedges’ g, and linear discriminant analysis (LDA) were used to characterise inter-species differences. H. ovalis was clearly distinguished from both co-occurring species: the Hd. uninervis–H. ovalis pair showed a discriminating window of 692–1394 nm (mean |g| = 1.31, BH q = 0.000046), and that for the H. stipulacea–H. ovalis pair was 700–1376 nm (mean |g| = 1.21, BH q = 0.000285); the JM distances were 1.60–1.67. A secondary shortwave-infrared discriminating window (1607–1755 nm; mean |g| = 0.90, BH q = 0.006) was also identified for the Hd. uninervis–H. ovalis pair. The H. stipulacea–Hd. uninervis pair showed meaningful geometric separation (JM = 0.994) but no individually significant wavelengths at the available sample size. ASentinel-2-proxy LDA achieved 85.6% overall accuracy (balanced accuracy = 87.3%; macro area under the curve = 0.917), outperforming a Landsat-proxy model by 20 percentage points. For each species, both a best-overall index and a visible-range alternative optimised for submerged satellite remote sensing are reported. The primary indices achieved balanced accuracies of 0.877–0.924; the visible-range alternatives achieved 0.818–0.907. Performance degraded substantially under noise (σ ≥ 0.002: −7.5 percentage points [pp]) and wavelength misregistration (±2–3 nm shifts caused losses of 5.5–15.7 pp), calling for stringent calibration requirements. These results constitute the first multi-species spectral library for Kingdom of Bahrain seagrasses, supporting Sentinel-2-based species mapping in the Arabian Gulf.

1. Introduction

Seagrass meadows are among the most ecologically and biogeochemically significant coastal ecosystems [1,2]. They represent globally significant organic carbon stocks [1], stabilise sediments, improve water quality, and provide essential nursery habitats for commercially important fish and endangered megafauna, including dugongs and green sea turtles [2]. Waycott et al. [3] estimated that seagrass meadows are declining globally at a rate of approximately 7% per year—a rate comparable to that for mangroves, coral reefs, and tropical rainforests [3]. Accurate monitoring of seagrass distribution, extent, and species composition is therefore a conservation priority. However, despite growing knowledge of global seagrass distribution and extent [4], mapping species composition from satellite imagery at an ecologically meaningful spatial resolution remains a largely unsolved challenge [4,5,6,7]. Beyond the Arabian Gulf, recent global syntheses summarise progress and outstanding methodological gaps in seagrass remote sensing [8,9], with species-level mapping demonstrated at scale in Mediterranean systems using time-series satellite analysis [10], in Australian shallow-water systems combining field validation with object-based image analysis [11], and in Florida using airborne hyperspectral data [12]. Closer to home, recent unmanned aerial vehicle (UAV) and Sentinel-2 benthic-habitat mapping in the NEOM coastal waters of the northern Red Sea has demonstrated the operational feasibility of optical satellite seagrass detection in adjacent hypersaline waters [13]; species-level discrimination, however, remains a gap that is not addressed by the broad benthic classes typically used in such regional studies. Across these regions, the persistent gap remains species-level discrimination at scales relevant to coastal management, particularly for taxa whose narrow-band spectral signatures have not been systematically characterised against operational satellite sensors.

Species-level discrimination based on remote sensing requires preserving narrow spectral absorption features—typically 5–20 nm wide—that are diagnostic of differences in photosynthetic pigment composition and leaf structure among taxa [5,14]. Seagrass leaves absorb strongly in the blue (400–500 nm) and red (650–700 nm) wavelengths owing to chlorophyll a, chlorophyll b, and carotenoids, with reflectance peaks in the green range (520–580 nm) and a sharp red-edge rise at 690–720 nm [5,14]. Distinguishing seagrass species optically is complicated by water column attenuation at depth, epiphyte fouling, substrate mixing, and intraspecific variability that can exceed interspecific differences in parts of the spectrum [5,14].

The Arabian Gulf is a biogeographically extreme province with sea surface temperatures ranging from 15 to 35 °C seasonally, salinities consistently above 40 psu, shallow bathymetry, and variable turbidity [15]. Seagrass meadows in the Gulf Cooperation Council (GCC) countries on the western and southern shores of the Gulf—Kingdom of Bahrain, Saudi Arabia, Kuwait, Qatar, UAE, and Oman—provide a critical habitat for dugongs (Dugong dugon), green sea turtles, and economically important fish but face intense pressure from coastal development, dredging, and climate change [16,17,18]. Despite these ecological stakes, data on the spectral characterisation of GCC seagrasses are virtually absent in the literature. A structured search of the peer-reviewed literature identified only two studies reporting field spectral measurements from the Kingdom of Bahrain [19,20]; a Sentinel-2 habitat-mapping effort in the UAE [21] represents the only comparable effort in an adjacent GCC state. No published studies have included Halophila ovalis in a Kingdom of Bahrain spectral analysis, and no full-range (350–2500 nm) spectral libraries have been reported for any GCC seagrass species.

This study addresses these gaps by (i) establishing full-range (350–2500 nm) spectral signatures for all three dominant Kingdom of Bahrain seagrass species; (ii) identifying statistically significant discriminating wavelengths using non-parametric tests with multiple-comparison correction; (iii) quantifying inter-species separability using geometric and probabilistic distance metrics; (iv) designing and validating six species-specific vegetation indices (two per species); (v) evaluating a Sentinel-2-proxy lineardiscriminant analysis (LDA) classifier; and (vi) assessing classifiers’ sensitivity to measurement noise and wavelength misregistration.

2. Materials and Methods

2.1. Study Area and Species

Field sampling was conducted in the coastal waters of Bahrain Island in the Arabian Gulf from May to October 2025, encompassing the peak summer growing season in Bahrain’s coastal waters. The three target species—Halophila stipulacea (Forsskål) Ascherson, Halodule uninervis (Forsskål) Ascherson, and Halophila ovalis (R.Br.) Hook.f.—are the dominant seagrass taxa documented in Kingdom of Bahrain waters [16], and they are widely distributed across the Indian Ocean and Indo-West Pacific region [2]. Sampling covered 29 geographic stations (52 species–station sampling units in total) distributed across the shallow coastal waters around Bahrain Island (Figure S1).

At each station, 0.5 m × 0.5 m quadrats were surveyed, and leaf tissue samples were collected by divers. Leaf length was measured from approximately ten randomly selected leaves per species per station using a calibrated ruler. The collected material was sealed in labelled sample bags and transported on ice to the laboratory for same-day spectral measurement. The sampling and measurement workflow is documented in Figure 1. In the following, we use HS, HU, and HO as shorthand for H. stipulacea, Hd. uninervis, and H. ovalis, respectively.

2.2. Spectral Data Collection

Spectral reflectance was measured in the laboratory using an ASD FieldSpec 4 Hi-Res spectroradiometer (Malvern Panalytical, Malvern, UK; spectral range—350–2500 nm; sampling interval—1 nm; spectral resolution—3 nm at 700 nm, and 8 nm at 1400 nm and 2100 nm). The hardware scan time was fixed at 100 ms across the full visible–near-infrared (VNIR)–shortwave-infrared (SWIR1)–SWIR2 detector range, and the base integration time was automatically optimised by the ASD Spectrometer App (Malvern Panalytical, Malvern, UK) to suit laboratory lighting conditions at the start of each measurement session. The instrument was mounted on a laboratory goniometer (a circular arc rail system, custom-built at the Arabian Gulf University, Bahrain) that enables controlled multi-angular spectral measurements, ensuring illumination geometry is repeatable and minimising directional reflectance artefacts. The fibre-optic contact probe was positioned at the nadir relative to the sample surface at a fixed height, with a 25° field of view. The goniometer’s multi-angular ability was not utilised in this study; all measurements were acquired at a single, fixed nadir viewing angle. Measurements were referenced to a Spectralon^® white reference panel (Labsphere Inc., North Sutton, NH, USA) at the start of each session and after every ten spectra to account for lamp drift. Each spectrum comprised 10 co-averaged scans, as per the ASD FieldSpec 4 standard measurement protocol, and the raw spectra were exported using ViewSpecPro version 6.2.0 (Malvern Panalytical, Malvern, UK). Prior to measurement, each leaf sample was washed and gently blotted to remove epiphytes, debris, and surface seawater; the triplicate samples (Q1–Q3) collected at each station were sorted by species before measurement to ensure spectral purity. The samples were placed on a dark reference panel (a near-zero-reflectance surface) to eliminate substrate reflectance contributions (Figure 1b). No detector splice correction was applied at the VNIR–SWIR1 boundary, and no smoothing was applied to the raw spectra; the quality-control range (400–900 nm, Section 2.3) lies wholly within the VNIR detector and is therefore unaffected by splice artefacts. A total of 100 raw spectra were collected across 52 species–station sampling units: 47 for H. stipulacea (25 stations), 36 for Hd. uninervis (19 stations), and 17 for H. ovalis (8 stations). The prominence of above-water SWIR reflectance (>1000 nm) for all three species is consistent with controlled laboratory conditions, as water-column absorption eliminates SWIR signals at depths below approximately 1–2 m [22].

2.3. Quality Control

An automated quality control (QC) protocol excluded spectra containing negative reflectance values in the spectrally reliable wavelength range: 400–900 nm. This range was chosen because it encompasses the most reliable portion of the ASD FieldSpec^® 4 visible and near-infrared (VNIR) detector (350–1000 nm): the 350–400 nm region is excluded due to an inherently low signal-to-noise ratio at short wavelengths, and the 900–1000 nm region is excluded to avoid detector-junction splice artefacts that can produce apparent negative values at the VNIR–SWIR1 transition. Negative reflectance in the 400–900 nm range is physically unmeaningful and indicates a measurement failure (a missed white-reference panel, sample movement, or sensor malfunction). Three spectra were excluded—one pertaining to H. stipulacea and two corresponding to Hd. uninervis—yielding a final dataset of 97 spectra: 46 for H. stipulacea, 34 for Hd. uninervis, and 17 for H. ovalis. Retention rates were 98%, 94%, and 100%, respectively.

2.4. Statistical Analysis

Kruskal–Wallis (KW) H-tests were applied across the entire spectral range measured (350–2500 nm; 2151 wavelength bands) for the three-way species comparison [23]. Discriminating windows and uniqueness profiles were interpreted across the full 350–2500 nm range, including SWIR features. The KW test is a non-parametric rank-based analogue of one-way analysis of variance (ANOVA) that requires no assumption of normality. Its use here is justified on three grounds: (i) the per-wavelength group sizes (n = 17–46) are too small to reliably verify normality across all 2151 wavelengths; (ii) preliminary Shapiro–Wilk tests [24] indicated non-normal distributions at many wavelengths, particularly in the near-infrared (NIR) plateau where within-species variance is high; and (iii) the test is robust to the unequal sample sizes characteristic of opportunistic field sampling. Pairwise Mann–Whitney U tests were then conducted at every wavelength for each of the three species pairs [25] (Dunn’s test, which uses the same KW ranks, is a standard rank-based post hoc procedure for the KW test; Mann–Whitney U was used here because it yields directional per-wavelength effect sizes that align directly with Hedges’ g computation and facilitates interpretable spectral window reporting).

Within each pairwise species comparison, all 2151 per-wavelength Mann–Whitney p-values were independently adjusted using the Benjamini–Hochberg (BH) false discovery rate (FDR) procedure [26], yielding three separate sets of q-values. Correction was applied per comparison rather than globally across all 6453 tests because each pairwise contrast constitutes a distinct hypothesis family: the question of which wavelengths differentiate H. stipulacea from Hd. uninervis is biologically independent of the question of which wavelengths differentiate either species from H. ovalis, and the resulting discriminating windows are reported and interpreted separately for each pair. Bonferroni or other family-wise error rate (FWER) corrections would be excessively conservative for spectrally autocorrelated data, for which adjacent wavelengths carry correlated information and identifying clusters of significant wavelengths is more meaningful than conducting individual tests [26]. BH correction controls the expected proportion of false positives among all rejected hypotheses, constituting the appropriate criterion for discovery-oriented spectral screening. All Mann–Whitney U tests were two-tailed; the direction of inter-species reflectance differences was captured by the sign of Hedges’ g rather than by one-tailed p-values. ASD FieldSpec^® 4 measurements are continuous floating-point values with sufficient numerical precision to render tied ranks negligible across the 97 spectra, and no tie correction was applied. Effect sizes were quantified as Hedges’ g [27] (bias-corrected for small samples), with |g| ≥ 0.8 serving as the conventional large-effect threshold [27]; this cutoff was chosen to ensure that the reported discriminating windows reflect robust inter-species differences rather than statistically significant but radiometrically trivial reflectance variations. Consecutive wavelengths meeting both BH-corrected significance (q < 0.05) and |g| ≥ 0.8 were merged into contiguous discriminating spectral windows, requiring a minimum run of five consecutive 1 nm bands (≥5 nm) to exclude isolated significant wavelengths that may represent noise artefacts rather than genuine discriminating regions.

2.5. Separability Metrics and Uniqueness Score

Six pairwise separability metrics were computed per species pair: Euclidean distance in spectral feature space, spectral angle mapper (SAM; degrees) [28], mean absolute difference (MAD), root mean square error (RMSE), Bhattacharyya distance, and Jeffries–Matusita (JM) distance. Bhattacharyya distance is a classical information-theoretic measure of separability between class-conditional multivariate distributions; it was computed on principal component analysis (PCA)-reduced data to circumvent the singular covariance problem at n < p, retaining up to five principal components (n_components = min(5, n − 2, p), which resolves to five at all three pairwise sample sizes in this study. Laboratory hyperspectral reflectance spectra are highly collinear across adjacent wavelengths, and five components are sufficient to capture the dominant axes of spectral co-variation while remaining well within the rank constraint of each pairwise dataset (the most constrained being the HU–HO pair with combined n = 51). JM distance is also PCA-reduced and bounded 0–2; values > 1.5 were adopted as the threshold for moderate-to-good class separability, and values approaching 2.0 indicate near-complete distributional separation. JM distance is preferred here over Bhattacharyya distance because it saturates at 2.0, providing an interpretable upper bound and preventing artificially large values from inflating perceived separability.

A per-species uniqueness score was computed for each species s at each wavelength λ according to

U(λ, s) = min_j D(λ, s, j) × 1/(1 + CV_s(λ))

(1)

where the pairwise discrimination score D(λ, s, j) is defined as follows:

D(λ, s, j) = |μ_s(λ) − μ_j(λ)|/(σ_s(λ) + σ_j(λ))

(2)

This form is structurally analogous to the discrimination score employed by Kaufman and Remer [29], originally developed for detecting forest cover from mid-infrared reflectance in the context of aerosol correction. Here, j indexes all comparison species, and CV_s(λ) = σ_s(λ)/μ_s(λ) is the within-species coefficient of variation. Taking the minimum across pairwise comparisons requires each species to be simultaneously distinctive from all co-occurring taxa; the weight 1/(1 + CV) penalises wavelengths with high within-species variability. The per-wavelength separability framing follows the approach developed by Rowan et al. [30], who evaluated the spectral separability of submerged aquatic vegetation at each wavelength across scales to identify discriminating spectral regions. The combined minimum-discriminability plus stability-weighting formulation is unique to this study. The uniqueness score serves as a descriptive, diagnostic and visualisation tool: it identifies wavelength regions with the greatest per-species spectral distinctiveness and motivated the sensor assessment in Section 3.8. The pairwise discrimination scores D(λ, s, j) from which uniqueness is derived underpin the uniqueness score calculation. Feature selection for vegetation index design relied independently on ANOVA F-test ranking and cross-validation (CV)-validated classification performance (Section 2.6) as opposed to the uniqueness score directly.

2.6. Feature Selection and Classification

Species-specific vegetation indices (VIs) were designed based on a candidate pool of wavelengths shortlisted via ANOVA F-test ranking [31,32] (the top 24 wavelengths with ≥10 nm minimum spacing—the 24-candidate limit provides a computationally tractable pool spanning the full spectral range, while the 10 nm spacing meets or exceeds the instrument’s spectral resolution at all wavelengths (3 nm at 700 nm, 8 nm at 1400 nm), ensuring candidates are spectrally non-redundant), augmented with wavelengths corresponding to known seagrass absorption features and the discriminating windows identified in the KW and Hedges’ g analysis, following Pu et al.’s design philosophy [31,32]. For each species, binary classification performance (target species vs. all others, balanced accuracy and area under the curve (AUC) via repeated stratified 5-fold CV) was used to select the optimal VI formula from single-band, normalised-difference, and spectral-window candidates; normalised-difference formulas were preferred when a result was within a balanced accuracy of 0.03 with respect to the overall best (this 0.03 threshold is an original methodological choice adopted for this study to favour physically interpretable ND formulations over single-band or spectral-window formulas with marginally higher raw performance; it is not a published standard). Although selection was data-driven, the resulting candidate pool comprises wavelengths corresponding to known plant-physiological features: the red-edge transition (700–730 nm), chlorophyll-a red absorption maximum (665–680 nm), NIR plateau structure (730–865 nm), and the green reflectance peak (560 nm), providing physical interpretability for the selected band combinations.

Five classifier models were evaluated, namely, (i) a best single-band decision tree baseline; (ii) a custom pairwise-feature decision tree (depth = 2); and (iii) an adapted three-index LDA using the three conventional broadband vegetation indices most widely applied in vegetation remote sensing as a baseline—the Normalised Difference Vegetation Index (NDVI), Green Normalised Difference Vegetation Index (GNDVI), and Normalised Difference Red Edge index (NDRE) [33,34,35]:

NDVI = (R₈₀₀ − R₆₈₀)/(R₈₀₀ + R₆₈₀)

(3)

GNDVI = (R₈₀₀ − R₅₅₀)/(R₈₀₀ + R₅₅₀)

(4)

NDRE = (R₇₅₀ − R₇₀₅)/(R₇₅₀ + R₇₀₅)

(5)

(iv) A Sentinel-2-proxy four-index LDA using indices computed from 7 nm mean reflectance windows (±3 nm) centred on Sentinel-2 band-centre wavelengths (705, 740, 783, 665, and 560 nm):

ND(783, 560) = (R₇₈₃ − R₅₆₀)/(R₇₈₃ + R₅₆₀)

(6)

ND(740, 705) = (R₇₄₀ − R₇₀₅)/(R₇₄₀ + R₇₀₅)

(7)

ND(783, 705) = (R₇₈₃ − R₇₀₅)/(R₇₈₃ + R₇₀₅)

(8)

ND(705, 665) = (R₇₀₅ − R₆₆₅)/(R₇₀₅ + R₆₆₅)

(9)

(v) A Landsat-proxy three-index LDA using 11 nm windows (±5 nm) centred on Landsat Operational Land Imager (OLI) band centres (865, 655, and 561 nm):

ND(865, 655) = (R₈₆₅ − R₆₅₅)/(R₈₆₅ + R₆₅₅)

(10)

ND(865, 561) = (R₈₆₅ − R₅₆₁)/(R₈₆₅ + R₅₆₁)

(11)

RATIO(655, 561) = R₆₅₅/R₅₆₁

(12)

Band values are estimated as simple rectangular-window means from the 1 nm field spectra rather than full spectral response function convolution; this proxy approach is a reasonable approximation for evaluating relative band utility, but it will overestimate performance compared with actual satellite data, where the broader bandwidths (15–37 nm) would reduce spectral contrast [36,37]. LDA was selected as the primary classifier [38] because it is well-suited to small, moderately imbalanced datasets. LDA assumes within-class multivariate normality and equal covariance matrices across classes. These assumptions are supported here on three grounds: (i) laboratory acquisition under controlled illumination and fixed viewing geometry minimises extraneous sources of within-class spectral variance, making reflectance at diagnostic wavelengths approximately Gaussian within each species; (ii) the classifier operates on a small set of four pre-selected vegetation index values rather than raw high-dimensional spectra, substantially reducing the risk of assumption violations relative to full-spectrum LDA; and (iii) perfect directional consistency across all cross-validation folds (1.0 for all six species-specific indices [33,34,35], as noted in Section 3.6) indicates stable, well-separated class boundaries consistent with a linear discriminant model. LDA also provides an interpretable linear decision boundary, which enhances reproducibility and applicability to independent datasets. All models were evaluated using repeated stratified k-fold cross-validation [39] (k = 5 folds, 5 repeats, 25 total splits—stratified to preserve class proportions across folds), reducing the partition-dependent variance in performance estimates and ensuring accuracy estimates are unbiased with respect to the smallest class (H. ovalis, n = 17). Leave-one-out CV is an alternative frequently recommended for n < 100 datasets [39]; repeated stratified k-fold was preferred here because stratification ensures each fold contains representatives of all three classes, which the leave-one-out method cannot guarantee for the smallest class. Balanced accuracy [40] was used as the primary performance metric rather than overall accuracy, because the latter is inflated for minority classes in imbalanced datasets and defined as the arithmetic mean of per-class recall rates.

Sensitivity analysis was employed to test balanced accuracy under additive Gaussian noise (σ = 0.001–0.005), reflectance quantisation (steps 0.001–0.01), and wavelength shifts (±1 to ±3 nm), evaluating the classifiers’ robustness to realistic instrument calibration uncertainties. The noise range (σ = 0.001–0.005) spans the ASD FieldSpec^® 4 noise-equivalent reflectance (NEdR ≈ 0.001 in the VNIR, as per manufacturer specifications) up to a conservative five-fold margin that accounts for inter-session calibration drift and suboptimal white-reference conditions. The wavelength shift range (±1–±3 nm) brackets the ASD FieldSpec^® 4 wavelength accuracy specification (±1 nm VNIR, as per manufacturer specifications) and extends to the operational tolerance likely encountered when applying the indices to other spectrometers or to satellite sensors with imperfect band-centre registration. The complete data acquisition and processing pipeline is summarised in Figure 2.

3. Results

3.1. Dataset Overview

After quality control, the analysis dataset comprised 97 spectra from 52 species–station combinations: 46 H. stipulacea (25 stations) spectra, 34 Hd. uninervis (19 stations) spectra, and 17 H. ovalis spectra (8 stations).

Table 1. Dataset overview: sample and station counts, quality-control retention, and depth ranges by species.

Species	Code	n Spectra (Total)	n Retained	n Excluded	% Retained	n Stations
Halophila stipulacea	A	47	46	1	98%	25
Halodule uninervis	B	36	34	2	94%	19
Halophila ovalis	C	17	17	0	100%	8

summarises sample and station counts, QC retention rates, and depth ranges. No significant relationships between collection depth and key spectral features were detected within the sampled depth range (Figure S2). QC retention was high for all species (94–100%), indicating generally good measurement quality.

3.2. Mean Spectral Signatures

All three species show the canonical seagrass spectral form—blue–red absorption troughs flanking a green reflectance peak, a steep red-edge rise, and an NIR plateau—as illustrated in Figure 3 [5,14]. Figure 4 shows the ±1 standard deviation (SD) envelopes, revealing that H. stipulacea and Hd. uninervis show considerably wider NIR variability than H. ovalis.

The mean photosynthetically active radiation ((PAR) 400–700 nm) reflectance, computed from continuous spectra, was 7.01% for H. stipulacea, 7.20% for Hd. uninervis, and 6.52% for H. ovalis—values consistent with the range reported for seagrass species globally (typically 4–10%) [5,41]. The most pronounced inter-species divergence occurs across the NIR range (700–1100 nm), peaking near 1098 nm (maximum inter-species reflectance range = 0.097), where H. ovalis is markedly lower than both co-occurring taxa. At 740 nm, mean reflectance was 0.1425 (H. stipulacea), 0.1232 (Hd. uninervis), and 0.0877 (H. ovalis). Key wavelength reflectance values are tabulated in

Table 2. Mean reflectance (±SD) at key diagnostic wavelengths according to species.

Wavelength (nm)	Region	H. stipulacea (Mean ± SD)	Hd. uninervis (Mean ± SD)	H. ovalis (Mean ± SD)
443	Violet	0.0607 ± 0.0179	0.0616 ± 0.0135	0.0614 ± 0.0135
490	Blue	0.0636 ± 0.0190	0.0644 ± 0.0141	0.0629 ± 0.0144
531	Green	0.0750 ± 0.0226	0.0712 ± 0.0155	0.0665 ± 0.0140
560	Green	0.0775 ± 0.0234	0.0751 ± 0.0169	0.0674 ± 0.0141
620	Red	0.0738 ± 0.0219	0.0778 ± 0.0177	0.0666 ± 0.0166
665	Red	0.0680 ± 0.0191	0.0768 ± 0.0169	0.0657 ± 0.0180
705	Red edge	0.1078 ± 0.0361	0.1004 ± 0.0284	0.0768 ± 0.0178
740	Red edge	0.1425 ± 0.0560	0.1232 ± 0.0426	0.0877 ± 0.0195
783	NIR	0.1554 ± 0.0628	0.1366 ± 0.0494	0.0905 ± 0.0205
800	NIR	0.1596 ± 0.0650	0.1421 ± 0.0522	0.0911 ± 0.0207
865	NIR	0.1728 ± 0.0722	0.1614 ± 0.0630	0.0929 ± 0.0220

.

3.3. Inter-Species Separability

Table 3. Pairwise spectral separability metrics across 350–900 nm for all three species pairs.

Species Pair	n_A	n_B	Euclidean Dist.	SAM (°)	MAD	RMSE	Bhattacharyya (PCA)	JM Dist. (PCA)
H. stipulacea vs. Hd. uninervis	46	34	0.3481	3.9972	0.0052	0.0075	0.6875	0.9944
H. stipulacea vs. H. ovalis	46	17	1.7728	14.8641	0.0238	0.0383	1.6199	1.6041
Hd. uninervis vs. H. ovalis	34	17	1.8635	15.1583	0.0258	0.0403	1.7889	1.6657

presents pairwise separability metrics computed across the 350–900 nm range. The H. stipulacea-versus-Hd. uninervis pair shows modest geometric separation (Euclidean distance = 0.348, SAM = 4.0°, and JM = 0.994), indicating distributional overlap that falls below the JM > 1.5 threshold for reliable discriminability (Section 2.5). Both pairs involving H. ovalis show substantially higher separability—JM = 1.604 for H. stipulacea vs. H. ovalis, and JM = 1.666 for Hd. uninervis vs. H. ovalis—with Euclidean distances approximately five times larger than the HS–HU pair. JM values of 1.604 and 1.666 fall in the moderate-to-good class separability range (Section 2.5), evidencing the strong distributional separation between these pairs.

3.4. Statistically Significant Discriminating Windows

No single wavelength across the full spectral range (350–2500 nm) showed a statistically significant pairwise difference between H. stipulacea and Hd. uninervis after BH correction (lowest pairwise q = 0.688 in a candidate 648–688 nm window), indicating that the two species cannot be separated by any individual wavelength in the optical range at the present sample size.

Both pairs involving H. ovalis produced large, highly significant discriminating windows. For H. stipulacea vs. H. ovalis, the primary window spanned 700–1376 nm (mean |g| = 1.21, BH q = 0.000285). For Hd. uninervis vs. H. ovalis, the primary window covered 692–1394 nm (mean |g| = 1.31, BH q = 0.000046), with a secondary SWIR window at 1607–1755 nm (mean |g| = 0.90, BH q = 0.006). These windows are visualised in Figure 5 (Hedges’ g profiles) and Figure 6 (KW significance profiles).

Table 4. Discriminating spectral windows: start and end wavelengths, mean Hedges’ |g|, and minimum BH-corrected q-value for each species pair. † Windows listed for H. stipulacea vs. Hd. uninervis are candidate windows only; neither one meets the combined criteria for BH significance (q < 0.05) and large effect size (|g| ≥ 0.8) required for a confirmed discriminating window.

Species Pair	Window (nm)	Width (nm)	Mean |g|	Min q	Higher Reflectance
H. stipulacea vs. Hd. uninervis	648–688	41	0.473	0.688	Hd. uninervis
H. stipulacea vs. Hd. uninervis	721–732	12	0.384	0.981	H. stipulacea
H. stipulacea vs. H. ovalis	700–1376	677	1.206	0.000285	H. stipulacea
Hd. uninervis vs. H. ovalis	692–1394	703	1.313	4.6 × 10−5	Hd. uninervis
Hd. uninervis vs. H. ovalis	1607–1755	149	0.903	0.00602	Hd. uninervis

provides full window statistics.

3.5. Species Uniqueness and Characteristic Wavelengths

Uniqueness scores—which measure a species’ spectral distinctiveness relative to co-occurring species at each individual wavelength—peaked at different wavelength regions for each taxon (

Table 5. Top discriminating wavelengths per species, ranked by uniqueness score U(λ, s) (Section 2.4).

Species	Wavelength (nm)	Region	Uniqueness Score
Hd. uninervis	674	Red	0.2191
Hd. uninervis	677	Red	0.2189
Hd. uninervis	675	Red	0.2187
Hd. uninervis	679	Red	0.2186
Hd. uninervis	673	Red	0.2168
Hd. uninervis	678	Red	0.2159
Hd. uninervis	676	Red	0.2147
Hd. uninervis	680	Red	0.2123
H. ovalis	1038	SWIR	0.7141
H. ovalis	1059	SWIR	0.7138
H. ovalis	1058	SWIR	0.7133
H. ovalis	1051	SWIR	0.7132
H. ovalis	1073	SWIR	0.7131
H. ovalis	1074	SWIR	0.7129
H. ovalis	1050	SWIR	0.7129
H. ovalis	1000	SWIR	0.7124
H. stipulacea	721	Red edge	0.146
H. stipulacea	722	Red edge	0.146
H. stipulacea	723	Red edge	0.1458
H. stipulacea	725	Red edge	0.1452
H. stipulacea	724	Red edge	0.1452
H. stipulacea	729	Red edge	0.1452
H. stipulacea	731	Red edge	0.1449
H. stipulacea	728	Red edge	0.1447

; Figure 7). Hd. uninervis showed peak uniqueness at 674–680 nm (red band, uniqueness ≈ 0.22), where it exhibits relatively elevated red reflectance (>8th percentile of all wavelengths for that species). H. ovalis showed its highest uniqueness score at 1004–1143 nm (SWIR, uniqueness ≈ 0.71)—the largest value across all species-wavelength combinations in the dataset (>8th percentile). H. stipulacea peaked at 721–731 nm (red edge; uniqueness ≈ 0.15; >8th percentile). An important interpretive caveat applies: the uniqueness score is a single-wavelength metric and does not capture the discriminatory power of band-ratio indices, which exploit the spectral slope between two wavelengths rather than absolute reflectance at either one. As a result, the wavelengths selected for band-ratio vegetation indices by the ANOVA F-test and cross-validation procedure do not necessarily coincide with uniqueness peaks. For H. ovalis, the primary VI band (R₁₀₉₈). falls at the 96.8th uniqueness percentile—confirming convergent validity for this species, where single-band distinctiveness and classification performance agree. For Hd. uninervis and H. stipulacea, the best VIs exploit band-ratio contrasts (the NIR–SWIR slope for HU, and the green–red ratio for HS) rather than the single-wavelength peaks flagged by the uniqueness score; the VI wavelengths for these two species fall in the 8st–8nd percentile of their respective uniqueness distributions. This finding is consistent with the descriptive role of the uniqueness score and does not represent an inconsistency; the final VI selection was independently validated via cross-validated balanced accuracy (

Table 6. Species-specific vegetation indices: formulation, optimal threshold, balanced accuracy, and area under the curve (AUC).

Target Species	Index Formula	Submerged Applicable	Threshold	Bal. Accuracy	AUC	Dir. Consistency
Hd. uninervis	HU_VI = (R_865 − R_1008)/(R_865 + R_1008)	No	−0.0062	0.9241	0.9388	1.0
Hd. uninervis	HU_VI = (R_531 − R_680)/(R_531 + R_680)	Yes	0.0034	0.9071	0.9232	1.0
H. ovalis	HO_VI = R_1098	No	0.119	0.8768	0.8643	1.0
H. ovalis	HO_VI = mean(R_675:685) − mean(R_780:786)	Yes	−0.0344	0.818	0.8529	1.0
H. stipulacea	HS_VI = (R_560 − R_680)/(R_560 + R_680)	Yes	0.0302	0.8836	0.902	1.0

). Diagnostic spectral windows for Hd. uninervis, H. stipulacea, and H. ovalis are shown in Figure 8a, Figure 8b and Figure 8c, respectively.

3.6. Species-Specific Vegetation Indices

Six vegetation indices were designed and validated for three species, with a primary index (with the best overall performance) and visible-range alternatives (applicable in submerged remote sensing) determined for each taxon (Table 6, Figure 8). For Hd. uninervis, HU_VI is the best-performing primary index (balanced accuracy, 0.924; AUC, 0.939), exploiting an NIR–SWIR contrast:

HU_VI = (R₈₆₅ − R₁₀₀₈)/(R₈₆₅ + R₁₀₀₈)

(13)

For submerged satellite applications, HU_VI_su^b uses visible wavelengths (balanced accuracy, 0.907; AUC, 0.923):

HU_VI_su^b = (R₅₃₁ − R₆₈₀)/(R₅₃₁ + R₆₈₀)

(14)

The negative HU_VI_su^b values characteristic of Hd. uninervis (Figure 8) reflects its sparse, narrow-bladed canopy architecture relative to the broader-leaved co-occurring taxa (the corresponding mechanism is discussed in Section 4.3). For H. ovalis, HO_VI is the best-performing primary index (balanced accuracy, 0.877; AUC, 0.864), reflecting H. ovalis’s uniquely low SWIR reflectance:

HO_VI = R₁₀₉₈

(15)

HO_VI_su^b, a visible-range window alternative, achieves reduced but operationally useful performance (balanced accuracy, 0.818; AUC, 0.853):

HO_VI_su^b = mean(R_675–685) − mean(R_780–786)

(16)

For H. stipulacea, HS_VI, a green–red normalised difference index, achieved a balanced accuracy of 0.884 and an AUC of 0.902; it uses visible wavelengths and therefore serves both the primary and submerged-applicable tiers:

HS_VI = (R₅₆₀ − R₆₈₀)/(R₅₆₀ + R₆₈₀)

(17)

All indices showed perfect directional consistency (1.0) across cross-validation folds. The distribution of VI values for each species across all six indices is shown in Figure 9.

3.7. Classification Performance

The Sentinel-2-proxy four-index LDA classifier achieved the highest cross-validated performance among the five tested models (

Table 7. Repeated stratified k-fold cross-validated performance (k = 5, 5 repeats) for all five classifier models: two decision-tree baselines and three linear discriminant analysis (LDA) configurations. Metrics: overall accuracy, balanced accuracy, macro F1, and macro one-versus-rest AUC.

Rank	Model	Feature Family	Estimator	Accuracy	Bal. Accuracy	F1-Macro	AUC (Macro OvR)
1	Sentinel-2 proxy 4-index LDA	sentinel2_proxy_lda	LDA	0.8557	0.8734	0.8415	0.917
2	Best single exploratory feature	single_feature	DecisionTree	0.7835	0.7954	0.7647	0.8731
3	Adapted 3-index LDA	adapted_standard_lda	LDA	0.7835	0.7805	0.7581	0.8877
4	Custom pairwise feature tree	pairwise_tree	DecisionTree	0.732	0.7246	0.702	0.8265
5	Landsat proxy 3-index LDA	landsat_proxy_lda	LDA	0.732	0.6726	0.6726	0.8544

, Figure 10 and Figure 11), with an overall accuracy = 85.6%, a balanced accuracy = 87.3%, a macro F1 = 84.1%, and a macro one-versus-rest AUC = 0.917 (Table 7). Per-class receiver operating characteristic (ROC) curves are shown in Figure 11. Per-class precision was highest for H. stipulacea (0.950, with a recall of 0.826 and an F1 of 0.884) and Hd. uninervis (0.906, with a recall of 0.853 and an F1 of 0.879), while H. ovalis achieved high recall (0.941) but lower precision (0.640, with an F1 of 0.762), reflecting the smaller support size (n = 17). The Landsat-proxy LDA performed substantially worse (balanced accuracy = 67.3%).

3.8. Sensor Assessment

To evaluate the feasibility of satellite-based discrimination of the three Kingdom of Bahrain seagrass species, we mapped the per-species uniqueness profiles against the band configurations of Sentinel-2 MultiSpectral Instrument (MSI) and Landsat 8/9 OLI (Figure 12,

Table 8. Alignment of Sentinel-2 MSI and Landsat 8/9 OLI band centres with the five diagnostic wavelengths identified in this study. Offsets > 10 nm are flagged †.

Diagnostic Wavelength (nm)	Sensor	Nearest Band	Band Centre (nm)	Bandwidth (nm)	Offset from Target (nm)
560	Sentinel-2	B3 (Green)	560	35	0
560	Landsat 8/9	B3 (Green)	561	57	1
665	Sentinel-2	B4 (Red)	665	30	0
665	Landsat 8/9	B4 (Red)	655	37	10
705	Sentinel-2	B5 (Red-edge)	705	15	0
705	Landsat 8/9	B4 (Red)	655	37	50 †
740	Sentinel-2	B6 (Red-edge)	740	15	0
740	Landsat 8/9	B4 (Red)	655	37	85 †
783	Sentinel-2	B7 (NIR)	783	20	0
783	Landsat 8/9	B5 (NIR)	865	28	82 †

† The 10 nm offset threshold exceeds the ASD FieldSpec^® 4 Hi-Res spectral resolution at 1400 nm (8 nm; 3 nm at 700 nm), ensuring that flagged band offsets exceed the instrument’s native spectral discrimination limit.

). The uniqueness score at each wavelength quantifies a species’ spectral distinctness with respect to all co-occurring taxa simultaneously—specifically, it corresponds to the minimum pairwise discrimination score weighted by spectral stability (1/(1 + CV)), penalising wavelengths with high within-species variability. Two complementary 90th-percentile thresholds are applied. For species-specific window identification (Table 5), the threshold is computed independently per species from that species’ own uniqueness distribution across reliable wavelengths, identifying the top 10% of wavelengths where each taxon is most self-consistently distinctive. For the sensor band assessment (Figure 12), the threshold is computed from the cross-species mean uniqueness at each wavelength, yielding a single global threshold (=0.270 for this dataset) that simultaneously identifies spectral regions that are informative across all three taxa. Both cutoffs correspond to the concept of the ‘diagnostic spectral window’ established by Fyfe [5] for seagrass discrimination and are original threshold choices for this study. They use a proportion-based criterion that requires no assumption about the absolute uniqueness score scale. The sensor band ranking is robust to the global threshold: testing at the 85th- and 95th-percentile thresholds yields identical top-band rankings for both Sentinel-2 (B7) and Landsat (B5).

For Sentinel-2 MSI, Band B7 (centre = 783 nm; bandwidth = 20 nm) [42] achieves the highest mean uniqueness score within any single band, and the B7/B8 combination provides the highest two-band mean. The 740–800 nm window corresponds to the seagrass red-edge and NIR plateau, where inter-species reflectance differences are governed primarily by leaf cellular architecture rather than photosynthetic pigment concentrations [5] (as discussed in Section 4.1). The 740–800 nm window also benefits from transmitting more readily through very thin (<1 m) clear-water columns as opposed to longer SWIR wavelengths [22], albeit with substantially higher water attenuation than visible wavelengths [43,44].

Landsat 8/9 OLI shows the best performance in Band B5 (865 nm, bandwidth = 28 nm) [45], but its complete absence of coverage in the 700–740 nm red-edge region critically limits its diagnostic utility relative to Sentinel-2. The wavelength offset of 10–85 nm for red-edge features means that Landsat OLI cannot access the spectral inflection zone where the steepest inter-species divergence occurs [20]. These findings align with the study by Bannari et al. [20], who concluded that Sentinel-2 MSI is more effective than Landsat for Kingdom of Bahrain seagrass discrimination, presenting broader evidence that multispectral sensors with coarser spectral resolution are generally insufficient for species-level mapping of seagrass [6,7].

A critical practical constraint applies to all NIR-based diagnostics: water absorbs NIR radiation strongly, with attenuation coefficients that are several times higher at 750 nm than at 550 nm in typical coastal waters [22]. The NIR spectral features identified here are therefore most applicable to (i) above-water measurements of fresh or air-exposed samples (as in this study); (ii) very shallow intertidal beds during low tide; (iii) uncrewed aerial vehicle (UAV)-based hyperspectral imaging, minimising water column depth; and (iv) prospective high-resolution satellite observations under optimal water-clarity conditions. Applying these NIR- and SWIR-based diagnostics to operational satellites will additionally require robust atmospheric correction and water-column correction prior to spectral matching [22,44]. The visible-range vegetation indices (HU_VI_su^b, HO_VI_su^b, and HS_VI (Section 3.6)) circumvent this water-column constraint by operating entirely within wavelengths transmitted through seawater [22].

3.9. Sensitivity Analysis

Non-repeated 5-fold stratified cross-validation (StratifiedKFold, n_splits = 5, random_state = 42) was used in the sensitivity analysis, producing out-of-fold predictions that cover all 97 samples, yielding a baseline balanced accuracy of 82.4% for the unperturbed Sentinel-2-proxy LDA (

Table 9. Sensitivity analysis: The balanced accuracy of the Sentinel-2-proxy LDA under increasing levels of Gaussian noise (σ), a reflectance quantisation step, and wavelength misregistration (±nm). Changes are reported as percentage-point differences from the unperturbed baseline (82.4%) †.

Perturbation Type	Level	Bal. Accuracy	Δ from Baseline (pp)
Baseline (no perturbation)	—	0.8244	0.0
Gaussian noise (σ)	+0.001	0.8171	−0.72
Gaussian noise (σ)	+0.002	0.7489	−7.54
Gaussian noise (σ)	+0.005	0.6436	−18.07
Reflectance quantisation (step)	+0.001	0.8244	0.0
Reflectance quantisation (step)	+0.002	0.8073	−1.71
Reflectance quantisation (step)	+0.005	0.7417	−8.27
Reflectance quantisation (step)	+0.01	0.6066	−21.78
Wavelength shift (nm)	−3	0.6944	−13.0
Wavelength shift (nm)	−2	0.7694	−5.5
Wavelength shift (nm)	−1	0.8274	0.3
Wavelength shift (nm)	+1	0.8022	−2.22
Wavelength shift (nm)	+2	0.7187	−10.57
Wavelength shift (nm)	+3	0.6671	−15.73

† The unperturbed baseline (82.4%) is lower than the repeated k-fold estimate in Table 7 (87.3%) because Table 9 uses non-repeated 5-fold CV (5 splits vs. 25 splits in Table 7); both values are unbiased but represent different estimators of generalisation performance.

). This baseline is lower than the repeated k-fold estimate (87.3%; Section 3.7) because that estimate uses 25 evaluation splits (5 folds × 5 repeats) as opposed to the 5 used here, providing additional variance reduction; neither value is biased, but they represent estimators with different degrees of precision. The single-repeat protocol is appropriate for sensitivity analysis because the key quantity of interest is the change in performance relative to the unperturbed baseline, not the absolute accuracy value; a fixed random seed and CV scheme ensure that any degradation observed is due solely to the applied perturbation rather than re-sampling variance. All perturbation-induced changes are reported as differences (percentage points) from this 82.4% baseline. Performance was robust to small perturbations: adding Gaussian noise at σ = 0.001 reduced balanced accuracy by only 0.72 percentage points (pp), and wavelength shifts of ±1 nm produced changes of less than ±2.3 pp. Performance degraded markedly under moderate perturbations: σ = 0.002 noise reduced accuracy by 7.54 pp; ±2–±3 nm wavelength shifts caused direction-dependent losses of 5.5–15.7 pp (+2 nm: −10.6 pp; −2 nm: −5.5 pp (Table 9)); and reflectance quantisation at step 0.01 caused a 21.8 pp reduction. These findings indicate that high-quality spectral calibration is essential for operational application of these classifiers, with the red-edge bands (705–740 nm) being particularly sensitive to wavelength misregistration due to the steep reflectance gradient in this region.

To complement the noise- and wavelength-calibration sensitivity analyses above, a 1000-cycle bootstrap resampling analysis was performed to test whether the limited n for H. ovalis (17 spectra, 8 stations) compromises the robustness of the discrimination findings. Resampling was performed with replacement within species groups. Mean balanced accuracy across cycles ranged from 0.846 (HU_VIsuψ, HO_VIsuψ) to 0.931 (HU_VI), with 100% direction consistency for all six species-specific indices. For the two visible-range indices most reliant on the H. ovalis signal (HU_VIsuψ and HO_VIsuψ), 27 and 28 of 1000 cycles, respectively, fell below the 0.70 balanced-accuracy threshold; these underperforming cycles correspond to bootstrap resamples that drew degenerate H. ovalis subsets dominated by single stations. The primary indices, the Sentinel-2-proxy LDA, and the H. ovalis-against-others discrimination remain robust under resampling (≥97% of cycles above threshold), confirming that the small H. ovalis sample contributes uncertainty but does not invalidate the main findings.

3.10. Top Kruskal–Wallis Wavelengths

Table 10. Top-ranking Kruskal–Wallis wavelengths (three-way H-test, BH-corrected) with H-statistic, BH q-value, and spectral region. See Figure 6 for the full significance profile across the measured spectrum.

Wavelength (nm)	Region	KW H Statistic	BH-adj. q Value	−log10(q)
1339	SWIR	21.53	0.000111	3.96
1338	SWIR	21.61	0.000111	3.96
1337	SWIR	21.45	0.000111	3.96
1336	SWIR	21.59	0.000111	3.96
1335	SWIR	21.73	0.000111	3.96
1334	SWIR	21.85	0.000111	3.96
1333	SWIR	21.98	0.000111	3.96
1332	SWIR	21.9	0.000111	3.96
1331	SWIR	21.86	0.000111	3.96
1330	SWIR	21.83	0.000111	3.96
1329	SWIR	21.77	0.000111	3.96
1328	SWIR	21.7	0.000111	3.96
1327	SWIR	21.87	0.000111	3.96
1326	SWIR	21.82	0.000111	3.96
1325	SWIR	22.06	0.000111	3.96
1324	SWIR	22.07	0.000111	3.96
1323	SWIR	21.96	0.000111	3.96
1322	SWIR	21.9	0.000111	3.96
1321	SWIR	21.84	0.000111	3.96
1320	SWIR	21.79	0.000111	3.96

lists the top-ranked wavelengths from the three-way Kruskal–Wallis analysis (BH-corrected). The full significance profile across 350–2500 nm is shown in Figure 6, which makes the spectral distribution of discriminating wavelengths immediately visible. The highest-ranked wavelengths according to H-statistic cluster are in the NIR region at 1080–1098 nm (H ≈ 22.2–22.3, BH q = 0.000111), primarily because of H. ovalis’s uniquely low NIR reflectance. A secondary high-ranked cluster occurs in the SWIR at 1320–1339 nm (H ≈ 21.5–22.1, same minimum BH q = 0.000111). While this region is inaccessible to conventional satellite sensors because of strong water absorption, this information confirms that SWIR spectral features carry the largest absolute inter-species signal—a finding consistent with the SWIR-based custom indices described in Section 3.6. Within the satellite-accessible range (≤900 nm), the top-ranked wavelengths lie in the red-edge and NIR (700–800 nm) regions, confirming their importance for Sentinel-2-based discrimination (Figure 5). Future UAV-based hyperspectral surveys of very-shallow seagrass beds could exploit SWIR features for species mapping without the water-column attenuation constraint [22].

4. Discussion

4.1. Spectral Distinctiveness of Halophila ovalis

The most robust finding from this study is the clear spectral distinctiveness of H. ovalis from both co-occurring Kingdom of Bahrain seagrass species. Three independent lines of evidence converge on this conclusion: (i) the JM distances (1.604–1.666) in the moderate-to-good class separability range (>1.5; Section 2.5), indicating strong but not near-complete distributional separation; (ii) the highly significant discriminating regions—Hd. uninervis–H. ovalis (692–1394 nm; BH q = 0.000046) and H. stipulacea–H. ovalis (700–1376 nm; BH q = 0.000285)—that survive the stringent FDR correction applied across 2151 wavelengths; and (iii) the consistently large effect sizes (mean Hedges’ |g| = 1.21–1.31), lying well above the conventional 0.8 threshold for practically meaningful differences [27]. Together, these metrics confirm that H. ovalis is not only statistically distinguishable but also practically separable from both co-occurring Kingdom of Bahrain taxa across a broad spectral range.

The dominant spectral mechanism is the markedly lower red-edge and NIR reflectance in H. ovalis (22–48% lower at 700–900 nm depending on wavelength and comparison species). Fyfe [5] experimentally demonstrated that NIR reflectance in seagrasses is primarily governed by leaf cellular architecture—particularly the number and arrangement of cell layers in mesophyll—rather than photosynthetic pigment concentrations. Based on this framework, the lower NIR reflectance of H. ovalis can most plausibly be attributed to its small, paddle-shaped leaves with fewer mesophyll cell layers and thinner blade cross-sections relative to the broad oval leaves of H. stipulacea or the linear blades of Hd. uninervis [2]. Direct measurement of H. ovalis leaf anatomy and chlorophyll content in Kingdom of Bahrain specimens would be needed to confirm this structural interpretation unambiguously.

H. ovalis also showed the highest uniqueness scores in the SWIR range (1004–1143 nm; uniqueness ≈ 0.71), a finding that has not been documented for Arabian Gulf seagrasses. While water absorption limits practical satellite application of SWIR features in submerged conditions [22], these results are directly relevant for above-water spectroscopy of collected samples and for uncrewed aerial vehicle (UAV)-based hyperspectral imaging of very-shallow seagrass beds.

4.2. Limited Differentiation of H. stipulacea and Hd. uninervis

No individual wavelengths in the 350–900 nm range passed BH correction for the H. stipulacea–Hd. uninervis pairwise comparison (lowest adjusted q = 0.688). This null result is best interpreted cautiously: at the available sample size (n = 46 and 34), this study does not have enough power to detect the modest effect sizes expected for two co-occurring congeners with broadly similar leaf morphologies. A post hoc power analysis (two-sample t-test approximation, α = 0.05, with an effective threshold adjusted for 2151 BH-FDR tests) indicated that a per-wavelength |g| of 0.3 (between the small [0.2] and medium [0.5] effect thresholds [27]) would require n ≥ 80 per species to achieve 80% power. This finding is therefore a statement about statistical power at the current sample size, not a demonstration that the two species are spectrally identical; the JM distance of 0.994 confirms there is real distributional separation at the multivariate level. These results are consistent with the study by Bannari et al. [20], who similarly found that single-band multispectral approaches cannot reliably distinguish these two Kingdom of Bahrain species.

At 665 nm, Hd. uninervis shows marginally but consistently higher mean reflectance (0.0768 ± 0.0169) than H. stipulacea (0.0680 ± 0.0191), indicating slightly lower red-band chlorophyll absorption in Hd. uninervis. By analogy with the package effect documented for other seagrass taxa [46], this finding could reflect differences in total chlorophyll concentration per unit leaf area, leaf blade thickness, or cell-layer organisation between the two species. Direct biochemical measurement would be needed to confirm this interpretation. Regardless of the underlying mechanism, the result has a clear practical implication for satellite-based mapping in the GCC: classifiers using broad multispectral bands are likely to consistently conflate H. stipulacea and Hd. uninervis, and operational seagrass products in the region should therefore report these two species as a combined class or rely on narrow Sentinel-2 red-edge bands with the calibration tolerances reported in Section 3.9.

4.3. Vegetation Indices and Operational Applicability

The custom indices present two tiers of operational utility. The primary indices (HU_VI, HO_VI, and HS_VI) represent maximum discriminatory performance and are best suited to above-water or intertidal spectroscopy and UAV-based hyperspectral imaging. The visible-range alternatives (HU_VI_su^b, HO_VI_su^b, and HS_VI) are directly applicable to satellite sensors such as Sentinel-2 for mapping submerged seagrass beds. For H. stipulacea, HS_VI is both the best-performing and most submergence-applicable index. For Hd. uninervis, the visible-range alternative HU_VI_su^b [(R₅₃₁ − R₆₈₀)/(R₅₃₁ + R₆₈₀)] achieved a balanced accuracy of 0.907 using green–red contrast, a value only 0.017 below the SWIR primary. For H. ovalis, the visible-range alternative HO_VI_su^b [mean(R_675–685) − mean(R_780–786)] incurs a larger accuracy penalty (0.818 vs. 0.877), consistent with the finding that H. ovalis is most distinctively characterised by its uniquely low SWIR reflectance—a feature inaccessible to submerged satellite sensors but recoverable via the NIR red-edge window at reduced confidence. The negative HU_VI_su^b values arise from Hd. uninervis’s sparse, narrow-bladed canopy architecture: low leaf-area coverage exposes the underlying sandy substrate, elevating 680 nm reflectance above 531 nm and inverting the green–red contrast observed in the broader-leaved H. ovalis and H. stipulacea, a spectral pattern consistent with structurally driven reflectance differences documented by Fyfe [5] and canopy-density sensitivity modelled by Hedley et al. [47].

4.4. Classifier Performance and Sentinel-2 Proxy

The Sentinel-2-proxy LDA significantly outperformed the Landsat-proxy model (balanced accuracy: 87.3% vs. 67.3%), confirming that Sentinel-2’s red-edge bands (705, 740, and 783 nm) provide meaningful discriminatory information on these species. This finding provides evidence of Sentinel-2’s potential utility in species-level seagrass mapping in the Kingdom of Bahrain, conditional on adequate water column correction. Clarke et al. [6] reported 85% genus-level accuracy with airborne hyperspectral imagery in temperate Australia, and Phinn et al. [7] demonstrated the advantage of fine-resolution multispectral sensors over broader-band alternatives for seagrass species discrimination, consistent with our finding that Sentinel-2’s 15 nm red-edge bands (705 and 740 nm [42]) provide discriminatory information in a spectral zone (700–740 nm) that Landsat OLI cannot access; its nearest band (B4: ~654 nm, 37 nm bandwidth [45]) falls ~50–86 nm short of the diagnostic red-edge inflection zone. Our comparison is based on convolving field spectra to simulated satellite band-equivalent reflectance; actual satellite imagery will additionally incorporate atmospheric path radiance, water column attenuation, and mixed-pixel contributions, all of which would reduce discrimination relative to field-measured performance [36,37]. The operational case for Sentinel-2 is reinforced by independent recent work in a Western Australian marine protected area, where Sentinel-2-derived indices—specifically the Normalised Difference Aquatic Vegetation Index (NDAVI) and a depth-invariant index of the blue and green bands—supported quantitative seagrass mapping in service of ecosystem-based fisheries management [48]. Whereas that study used non-specific vegetation indices to map seagrass cover, the species-specific indices reported here extend the same Sentinel-2 band set toward taxon-level discrimination, suggesting that the sensor can be used not only for presence/absence and cover but also, when paired with a regional spectral library, for biodiversity-level inventories.

The results of the sensitivity analysis have important operational implications. A wavelength misregistration of ≥2 nm reduces balanced accuracy by more than 5 pp; at ≥3 nm, the loss exceeds 10 pp. Red-edge bands (705, 740 nm) are especially vulnerable because the reflectance gradient in this region is very steep [5]. These considerations reinforce the need for rigorous sensor spectral calibration and atmospheric correction before applying the proposed indices to satellite data [42].

4.5. Comparison with the Published Literature

The mean PAR reflectance of all three Kingdom of Bahrain species (6.5–7.2%) is consistent with published ranges for seagrass species globally [5,41], confirming that the measurement conditions and sample handling did not systematically bias reflectance levels. The red chlorophyll absorption minimum at ~670 nm between this dataset (673 nm for H. stipulacea and 672 nm for Hd. uninervis) and values reported for the Kingdom of Bahrain by Bannari et al. [20] is consistent, reflecting methodological cross-validation. Pu et al. [31] demonstrated that hyperspectral narrow-band data in the green–red visible range (518–701 nm) yielded substantially better results relative to broadband Landsat imagery for seagrass mapping in a Florida coastal context; the visible-range wavelengths used in HS_VI (560 and 680 nm) are broadly consistent with the effective spectral region identified in the cited study. Figure 13 and

Table 11. Research comparison: key spectral features from this study versus published values for comparable seagrass species and spectral regions.

Species	Metric/Feature	Published Value	This Study	Direct Comparison	Reference
Halophila stipulacea; Halodule uninervis	Blue-absorption-feature position (continuum-removed reflectance)	485–498 nm	HS blue minimum, 452 nm; HU blue minimum, 452 nm	No	Bannari et al. [20]
Halophila stipulacea; Halodule uninervis	Green-reflection peak position (continuum-removed reflectance)	HS ~530 nm; HU ~544 nm	HS green peak, 559 nm; HU green peak, 580 nm	No	Bannari et al. [20]
Halophila stipulacea; Halodule uninervis	Red chlorophyll absorption minimum	~670 nm	HS red minimum, 673 nm; HU red minimum, 672 nm	Yes	Bannari et al. [20]
Halophila stipulacea; Halodule uninervis	Maximum inter-species reflectance difference under very dense submerged cover	<=6% visible; <=13% NIR	HS vs. HU maximum mean difference in this dataset: 0.96% across 400–700 nm and 1.98% across 701–900 nm	No	Bannari et al. [20]
Halophila ovalis within Australian seagrass comparison	Spectral zones with greatest between-species differences	500–600 nm and 700–750 nm	H. ovalis corresponds to the strongest unique window here = 1004–1143 nm; H. ovalis green peak = 558 nm	No	Durako [41]
Halophila ovalis	Photosynthetic leaf absorptance AL (PAR)	45 ± 3% to 62 ± 5%; mean 53 ± 5%	Not directly comparable from this dataset because transmittance was not measured, so AL(PAR) cannot be derived conservatively.	No	Durako [41]
Seagrasses in general	Leaf reflectance of PAR RL (PAR)	4.6–9.3%; mean 6.1 ± 1.2%	Mean reflectance over 400–700 nm: HS, 7.01%; HU, 7.20%; HO, 6.52%	Yes	Durako [41]
Seagrasses in general	Wavelength zones most affected by optical packaging	Green 500–600 nm most responsive; package effect strongest in blue 400–500 nm and red 600–700 nm	This dataset shows strong pigment-related structure in blue and red absorptions, with remaining visible separation concentrated in green-to-red windows and stronger species separation beyond the red edge.	No	Cummings and Zimmerman [46]
Seagrasses in general	Photosynthetic light-harvesting efficiency	approximately 50% of incident PAR	Not directly comparable because photon flux absorption efficiency requires radiometric weighting and absorptance rather than reflectance alone.	No	Cummings and Zimmerman [46]
Seagrass species discrimination	Optimal VI central wavelengths for species discrimination	460, 500, 610, 640, 660, 690 nm	Top compact discriminators in this dataset also draw heavily from visible pigment-sensitive bands, especially green and red contrasts such as 531/665 nm.	No	Pu et al. [31]

summarise quantitative comparisons with the available literature. The species-pair pattern observed here—high spectral separability of one taxon against two co-occurring congeners and limited separability between morphologically distinct species—is consistent with reports from Mediterranean Posidonia–Cymodocea–Caulerpa assemblages [10] and Australian Halophila–Halodule–Zostera assemblages [11], where similar asymmetries in species-pair discriminability have been documented. The GCC pattern reported here extends this body of work to a hypersaline, thermally extreme province whose seagrass spectral signatures had not previously been characterised across the full 350–2500 nm range. A complementary methodological line of work has reached species-level discrimination by replacing satellite multispectral imagery with high-resolution UAV imagery and deep-learning semantic segmentation: Tahara et al. [49] separated Zostera marina and Z. japonica in a Japanese lagoon at 82% overall accuracy. The Sentinel-2-proxy results reported here suggest that the two paths—field-derived spectral libraries feeding satellite indices, and UAV imagery feeding deep classifiers—are likely to be complementary rather than competing, with satellites better suited to regional inventories and UAV/deep-learning approaches better suited to fine-scale verification.

4.6. Limitations

Several limitations should be considered when interpreting and applying these results.

• Measurement geometry: All spectra were collected above water in the laboratory from freshly harvested samples. Leaf moisture state at the time of spectral measurement (e.g., whether the leaf surface retained a thin water film from ice-transport) was not systematically documented per spectrum. The presence of a surface water film may influence reflectance, particularly in the SWIR region, and limits direct inter-study comparison.
• Sample size imbalance: H. ovalis is represented by only 17 retained spectra from eight stations, limiting the precision of confidence intervals and contributing to lower classification precision (0.640) despite high recall (0.941). The bootstrap analysis reported in Section 3.9 confirms that the primary findings are robust to resampling, but caution is warranted when generalising the H. ovalis-specific thresholds or its visible-range index (HO_VIsuψ) to populations or seasons outside the sampled set, where the species’ spectral variability may be wider than captured here.
• Temporal snapshot: All measurements were collected from May to October 2025. Seasonal variation in chlorophyll content, canopy density, and leaf morphology may substantially alter spectral signatures [5]; the spectral library should not be treated as a phenological characterisation.
• Epiphyte status: Fouling level was not systematically documented; within-species spectral variability may partly reflect epiphyte load variation rather than a true species-specific signal [5].
• Lack of independent spatial validation: All accuracy metrics derive from cross-validated field spectra collected over one summer. Transferability to satellite imagery, different seasons, or spatially separated sites has not been tested and represents the necessary next step before operational application.
• SWIR applicability: The primary HU_VI and HO_VI use NIR–SWIR wavelengths that are inaccessible in submerged conditions [22] (see Section 4.3 for the two-tier operational framework and the performance of the visible-range alternatives). Future UAV-based hyperspectral surveys of very shallow seagrass beds could exploit SWIR features for species mapping without the water-column attenuation constraint [22]. Beyond SWIR inaccessibility, operational satellite mapping of submerged meadows also depends on water-column correction. Comparative work has shown that the standard depth-invariant and Sagawa correction models trade off accuracy against water clarity and depth, and that none performs universally well across complex shallow systems [52]; the choice of correction therefore needs to be calibrated against site-specific bathymetry and turbidity rather than applied as a default.
• Geographic scope: Sampling was confined to the coastal waters of Bahrain Island. Whether the spectral signatures and vegetation index thresholds are transferable to seagrass beds in other parts of the Kingdom of Bahrain or across the wider Arabian Gulf remains to be established with independent data.

5. Conclusions

This study provides the first quantitative multi-species spectral baseline for the three dominant seagrass species in the Kingdom of Bahrain, lying in the Arabian Gulf. The principal findings are summarised below.

Three independent lines of evidence establish Halophila ovalis as the most spectrally distinctive of the three Kingdom of Bahrain seagrass taxa: Jeffries–Matusita distances of 1.60–1.67 (moderate-to-good separability), broad discriminating windows spanning the red-edge and NIR regions (692–1394 nm versus Hd. uninervis; 700–1376 nm versus H. stipulacea), and large mean Hedges’ |g| values of 1.21–1.31 throughout these windows, all with BH q < 0.0003. In contrast, H. stipulacea and Hd. uninervis cannot be separated by any individual wavelength after BH-FDR correction at the available sample size, but a multi-band Sentinel-2-proxy LDA (balanced accuracy 87.3%) and the visible-range index HS_VI (AUC = 0.90) achieve practically useful discrimination.

Six species-specific vegetation indices are defined across two operational tiers: primary indices (HU_VI 0.924, HO_VI 0.877, HS_VI 0.884 balanced accuracy) deliver maximum performance for above-water and UAV surveys, while visible-range submerged-applicable alternatives (HU_VIsuψ 0.907, HO_VIsuψ 0.818, and HS_VI 0.884) enable satellite-based detection via Sentinel-2 MSI. Sentinel-2 MSI bands outperformed Landsat OLI by 20 percentage points in balanced accuracy, supporting targeted use of Sentinel-2 red-edge bands for Arabian Gulf seagrass mapping. Calibration requirements, however, are stringent—accuracy degrades by more than 7 pp at noise σ ≈ 0.002, and wavelength misregistration of ±2–3 nm causes losses of 5.5–15.7 pp—emphasising the need for well-characterised instruments.

Beyond these calibration constraints, and notwithstanding the geographic confinement of the dataset to the coastal waters of Bahrain Island and to a single summer growing season, the methodological framework—full-range field spectroscopy, separability metrics, species-specific vegetation index design, and sensor calibration sensitivity analysis—is directly applicable to other GCC and shallow-water marginal seagrass environments where comparable spectral libraries do not yet exist.