Influence of Environmental Factors on the Surface Feeding Behaviour of Immature Male Whale Sharks in the Gulf of Tadjoura (Djibouti)

Abstract

The East African country of Djibouti is known to host an important seasonal feeding aggregation of whale sharks that allows the frequent observation of their surface feeding behaviour. The influence of environmental factors on the different whale shark feeding strategies (passive, active, and vertical) was studied over a four-year period (2017, 2020, 2022, 2024) in the Gulf of Tadjoura. Across 81 immature male whale sharks identified and 1082 surface feeding behaviours recorded in this period, the chlorophyll-a concentration was the main parameter predicting the choice of the filter-feeding technique. Active and vertical feeding behaviours were associated with rainfall, lower sea surface temperature, worse sea conditions, and low wind speed during the morning, all factors positively correlated to chlorophyll-a concentration. On the contrary, passive feeding behaviour was favoured in the inverse environmental conditions. Both passive and vertical feeding behaviours occurred during El Niño events, whereas active feeding was more common during La Niña events. Since it is known that whale shark abundance and distribution are associated with food availability at coastal locations, it is fundamental to understand environmental drivers of filter-feeding strategies when managing conservation efforts for this endangered species. Recommendations for future research work at this site are presented.

1. Introduction

The whale shark (Rhincodon typus, Smith, (1828)) is the world’s largest filter-feeding fish living in the tropical oceans [1] and is known to form seasonal feeding aggregations in different sites of the world, including Djibouti [2,3,4,5,6]. Boldrocchi et al. [6] stated that whale sharks aggregate annually in the Gulf of Tadjoura (Djibouti) between October and February and the aggregation is principally made up of juvenile males. These aggregations allowed the development of the ecotourism industry and, consequently, the frequent observation of the whale shark surface feeding behaviour [7].

Since 1990, ecotourism of whale sharks has rapidly increased, resulting in positive impacts on local economies, conservation efforts, and public awareness of marine biodiversity when managed carefully to minimise negative effects on the animal and its feeding behaviour [3]. However, the frequent disregard of the code of conduct by tourists has led to a decrease in the number of R. typus present at the aggregation sites. These interactions can have effects both in the short and long term. Short-term effects observed in Ningaloo Reef (Western Australia) include barking, rapid diving, and avoidance of surface feeding behaviours [8]. On the other hand, long-term effects may include stress, avoidance, or displacement [1]. These observations notably contrast with the surface feeding behaviours observed in whale sharks in Oslob, Philippines, where sharks are provisioned daily and are less likely to exhibit avoidance, indicating an ability to modify their behaviour in response to anthropogenic stimuli [9].

During the seasonal aggregation, whale sharks in Djibouti feed on dense patches of zooplanktonic prey, consisting primarily of copepods, which account for more than 80% [6,10,11]. Boldrocchi et al. [6] also found a positive relationship between monthly chlorophyll-a concentration and zooplankton biomass in Djibouti, with an increase in the latter from October to its maximum peak in December. This is due to the southwest monsoon winds moving eastward and enhancing upwelling phenomena in the Gulf of Tadjoura [12]. The opposite trend occurs during winter, with northeast monsoon winds preventing the upwelling phenomenon, causing a decrease in chlorophyll-a and a subsequent decline in zooplankton abundance [6]. Moreover, Boldrocchi and Bettinetti [13] observed seven juvenile whale sharks feeding on a school of baitfish (anchovies) during the off-season, allowing them to meet their energetic demands when dense zooplankton patches were not available.

Different surface feeding behaviours have been described by several authors [1,14,15,16,17,18,19,20,21]. The most common ones are those reported by Nelson and Eckert [22], who highlighted three whale shark filter-feeding techniques exhibited in relation to specific zooplankton sources, resulting in an energy gain greater than the actual energy cost following the theory of optimal foraging. Ram (or simply passive) feeding has been mainly described when the zooplankton density is low, where whale sharks swim slowly in circular or S-shaped patterns just below the surface and pass water into a half-way- or wide-open mouth through their gills without the additional exertion of gulping or suction feeding. Vertical (suction or simply vertical) feeding is reported to be used when the zooplankton density is moderate. The whale shark body position is almost vertical with little or no forward movement with the mouth directed toward the surface. Water is forced to pass through gill slits by gulping and expelling water using a suction technique. Active surface ram (or simply active) feeding is employed when the zooplankton density is high. Sharks feed directly at the surface, swimming in a forward motion using both suction and ram feeding techniques to catch the prey. The upper jaw, the head, the first dorsal fin, and the upper caudal fin penetrate the sea surface and swimming patterns are frequently adjusted accordingly.

The correlation between filter-feeding techniques and zooplankton prey abundance was also observed by Di Capua et al. [11] in Djibouti. Indeed, passive feeding activity was recorded when zooplankton density was the lowest, while general suction feeding activities, comprising both vertical and active, were observed when high zooplankton density was recorded.

In addition to zooplankton density and anthropogenic factor, there is also growing concern that climate change may already be impacting whale sharks’ feeding behaviours [23] due to indirect effects such as changes to their habitat, food availability and densities, and ecosystem dynamics, which all represent a significant challenge to their survival and conservation efforts. It has already been observed that climate change can alter the distribution and productivity of planktonic prey, potentially affecting the behaviour and migration patterns of whale sharks and leading to food shortages in feeding grounds [24]. Based on current climate change trends, appearance of new suitable aggregation sites for whale sharks in some geographical area is expected, but suitable zooplankton productivity within these regions is not guaranteed. Conversely, a future decrease in the occurrence of zooplankton blooms in the current aggregation areas is hypothesised. Both climate-related events have the potential to negatively affect the feeding behaviour and migration patterns [25].

Several studies have described the influence of some environmental factors on whale shark movement patterns, sightings, and coastal aggregation [26,27,28,29,30]. However, overall effects of environmental factors on surface feeding behaviours and on the choice of the filter-feeding technique have never been described in the literature.

Motta et al. [19] observed in Cabo Catoche (Mexico) that whale sharks approached the surface to filter-feed during early morning, with a peak in abundance during mid-morning. Whale sharks returned to slightly deeper water around noon, resurfaced to feed in the afternoon, and returned to deep waters again in late afternoon, using ram filter-feeding techniques when swimming at a depth of between 0 and 1 m in the daytime. On the contrary, Gleiss et al. [31] stated that whale sharks at Ningaloo Reef exhibited ram filter-feeding techniques primarily during sunset and the first hours of the night, whereas vertical movement peaked earlier. Other predictors influencing oceanographic processes, such as the El Niño Southern Oscillation (ENSO) and related wind shear and rainfall, including temporal (year), oceanographic (SST, sea conditions), and large-scale factors, have been linked to shifts in the abundance and distribution of whale sharks in different places around the world [26,27,30,32,33] but not to the surface feeding behaviours of these animals. All these environmental factors can affect not only the whale shark’s distribution and abundance, but also its feeding behaviour and food resource availability, influencing the employment of whale shark filter-feeding techniques in relation to the density of their prey.

Given these important gaps and the threat posed to whale sharks at their coastal aggregation sites by shore-based fishing and boating activities, it is fundamental to understand drivers of their filter-feeding strategies when managing conservation efforts for this endangered species.

Thus, this study aims to (i) evaluate if whale shark surface feeding behaviour is potentially affected by environmental factors; (ii) investigate the main factors related to sea and weather conditions affecting the whale shark surface feeding behaviour; (iii) assess the existing relationship among environmental conditions; and (iv) evaluate in detail how environmental factors affected the recorded surface feeding behaviours.

2. Materials and Methods

2.1. Sampling Area

The study was performed along the coastal waters of Djibouti, in the Gulf of Tadjoura (11°40′ N, 43°00′ E), at the southern entrance to the Red Sea (Figure 1). This inlet of the Indian Ocean is the result of the fault line of the northerly end of the East African Rift Valley that transects Djibouti, Ethiopia, and Kenya [2]. Data were collected in two close feeding hot spots for whale sharks between Arta Beach (11°58′ N, 42°82′ E) and Ras Korali (11°34′ N, 42°47′ E), and fieldwork was conducted by the team from the Sharks Studies Center–Scientific Institute of Massa Marittima (GR, Italy) on board a sailing vessel called “Elegante”.

Sampling activities were performed in January 2017, January 2020, January 2022, November 2022, and January 2024. All sampling lasted 1 week except the one carried out in January 2022, which lasted two weeks.

2.2. Whale Shark Observations

Once the boat was anchored, two zodiacs were used to search for whale sharks in two 3 h surveys carried out between the following time intervals: 09.00–12.00 and 14.00–17.00. An average of 60 h each week (6 h for each zodiac per 5 days each week) and a total of 300 h of total observation time (60 h each year in 2017, 2020, and 2024; and 120 h in 2022) were collected over the course of the study. Observations randomly took place within 50 m from the coast in the two study areas.

Whale sharks were located by identifying the dorsal part of the head, dorsal fin, and/or the upper lobe of the caudal fin at the sea surface following the method used by Boldrocchi et al. [6]. Once a shark was spotted, research snorkelers approached it underwater and began taking photos and videos with action cams of both flanks above the pectoral fins, just posterior to the gill slits, to photo-identify the specimen. Body scars and wounds were also filmed, since they vary in both shape and size. They can also remain as permanent marks, allowing easy visual identification when photos of dorsal, pectoral, or caudal fin are missing [22]. Sex was determined by swimming underneath the shark and observing whether claspers were present on the inside of the pelvic fins, indicating a male shark [6,22]. During the approach to the shark, size was initially measured with a laser-photogrammetry survey (from the mouth to the beginning of the first dorsal fin) [34]. The complete total length (TL) was then calculated using the equation proposed by Matsumoto et al. [35]. However, the use of the laser-photogrammetry survey for the size estimate occurred only in 2022 and 2024.

Photographs were then loaded in the I³S (Interactive Individual Identification System) Classic pattern recognition software v4.0 to evaluate matches among specimens [5,6,36,37]. For each sighted and filmed shark, information on the identification number, photos of both flanks, date and time of sighting, sighting area, sex (if available), TL (if available), videos and photo references, and position of scars and wounds were recorded for each specimen. If the same shark was re-sighted several times during the same or different years, only the first recorded sighting was considered. When a shark of known size was re-identified through the laser survey, a comparison of TL was carried out to assess any size increase over time.

Videos were also analysed to classify surface feeding behaviours exhibited by sharks. Only videos where both mouth and gills were clearly visible were considered for evaluation. Feeding activities were classified as passive (P), active (A), or vertical (V) according to parameters described by Nelson and Eckert [22]. The frequency of each behaviour was calculated as a percentage (%). It should be noted that it was possible for one specimen to be observed at various times throughout the day and that all its exhibited behaviours were recorded.

2.3. Environmental Data Collection

Environmental data were collected in 2017, 2020, 2022, and 2024 as follows:

(1)

Sea conditions were obtained from the windguru database “https://www.windguru.cz/4910 (accessed on 1 April 2024)” for “Djibouti (East Africa)” area and classified as “calm” (0–10 cm wave height), “slightly rough” (11–50 cm wave height), and “rough” (>50 cm wave height).

(2)

Light levels were expressed in oktas, a unit of measurement that indicates the cloudiness of the sky, estimated in terms of how many eighths of it are obscured by clouds [38]. Measurement intervals used to assess the sky coverage were as follows:

(a)

0–2 oktas corresponded to clear sky.

(b)

3–5 oktas corresponded to a partly cloudy sky.

(c)

6–8 oktas corresponded to an overcast sky.

(1)

Sea surface temperature (SST), expressed in degrees Celsius (°C), was obtained by the underwater computer.

(2)

Wind speed, expressed in knots (Km/h), was obtained from the windguru database for the “Djibouti (East Africa)” area.

(3)

Rainfall, expressed in millimetres per hour (mm/h), was obtained from the windguru database for the “Djibouti (East Africa)” area.

(4)

Chlorophyll-a concentration, expressed in milligrams per cubic meter (mg/m³), was obtained from the Copernicus database https://data.marine.copernicus.eu/viewer/expert?view=viewer&crs=epsg%3A4326&t=1673136000000&z=0¢er=43.27123058951001%2C8.0075818162499&zoom=13.011481121057063&layers=W3siaWQiOiJjOCIsIm9wYWNpdHkiOjEsImxheWVySWQiOiJHTE9CQUxfTVVMVElZRUFSX0JHQ18wMDFfMDI5L2NtZW1zX21vZF9nbG9fYmdjX215XzAuMjVfUDFELW1fMjAyMTEyL2NobCIsInpJbmRleCI6MTAsImxvZ1NjYWxlIjpmYWxzZX0seyJpZCI6ImM5Iiwib3BhY2l0eSI6MSwibGF5ZXJJZCI6IkdMT0JBTF9BTkFMWVNJU0ZPUkVDQVNUX0JHQ18wMDFfMDI4L2NtZW1zX21vZF9nbG9fYmdjLXBmdF9hbmZjXzAuMjVkZWdfUDFELW1fMjAyMzExL2NobCIsInpJbmRleCI6MCwibG9nU2NhbGUiOmZhbHNlLCJjb2xvcm1hcElkIjoiYWxnYWUifV0%3D&basemap=dark (accessed on 1 April 2024) for Arta Beach and Ras Korali areas.”

(5)

Time of the day, expressed in hours (h), was obtained by the daily surveys in Arta Beach and Ras Korali areas following these scheduled times: 09.00–12.00 and 14.00–17.00.

(6)

El Niño Southern Oscillation (ENSO), expressed in terms of the Multivariate ENSO index (MEI), was obtained from the NOAA Climate Prediction Center database “https://origin.cpc.ncep.noaa.gov (accessed on 1 April 2024)”.

2.4. Empirical and Statistical Analysis

The dataset analysed in this study contains both indirect and unobserved factors to address endogeneity issues. The former corresponds to six environmental factors: (i) light levels, (ii) sea surface temperature (SST), (iii) rainfall, (iv) wind speed, (v) ENSO measurement unit (MEI), and (vi) chlorophyll-a concentration. Indirectly measured variables were computed as proxy discrete variables and consist of two additional environmental factors: (i) sea conditions in terms of activity as an ordinal variable, denoted as 1 (calm), 2 (slightly rough), and 3 (rough); and (ii) time periods during which sightings occurred as dummy variables equal to 0 (9:00–12:00) and 1 (14:00–17:00).

The dataset was further arranged to better analyse the degree of interdependence and/or relationship between indirect and unobserved factors. More precisely, some variables were grouped into classes and then evaluated as categorical or ordinal discrete indicators: SST, assuming values 0 for the class $\le$26 °C and 1 for the class $>$26 °C; rainfall, assuming values 0 for the class 0.0 mm/h and 1 for the class > 0.0 mm/h; wind speed, taking values 1 for the class 3.0–6.9 knots, 2 for the class 7.0–9.9 knots, and 3 for the class 10.0–12.9 knots; the MEI, assuming values 0 for the class between −0.9 and 0.0, and 1 for the class between 0.1 and 2.0; and presence of chlorophyll-a, taking values 1 for the class 0.00–0.50 mg/m³, 2 for the class 0.51–2.00 mg/m³, and 3 for the class $\ge$2.01 mg/m³. Every group of classes was computed based on its median relative to the sample size (1082 surface feeding behaviours recorded), so that each class was equally distributed and weighted when making statistical inferences.

The variable of interest “dbehaviour” corresponds to the whale shark surface feeding behaviour. It was computed as a dummy variable equal to 0 (passive) or 1 (active and vertical) for the first three research questions of this study and as an ordinal variable equal to 1 (active), 2 (passive), and 3 (vertical) for the fourth. Active and vertical behaviour were grouped together under the assumption that the whale shark is exhibiting active predation behaviour in both instances [11].

First, a chi-square test of independence was applied to investigate whether the surface feeding behaviours affected by environmental factors in two different sampling periods in 2022 (January and November) were independent (null hypothesis) or dependent (alternative hypothesis). More precisely, the hypothesis testing assumed that two variables (“dbehaviour” evaluated as a dummy variable and “months” as ordinal variables equal to 0 (January) and 1 (November)) were likely to be unrelated (independency under the null) or related (dependency under the alternative).

Before addressing the research questions, a further analysis was conducted to assess if the number of feeding behaviours displayed depended on the number of whale shark sightings. More precisely, a statistical test was performed to check for a significant association or correlation between these two variables. According to the data, both a regression analysis and Pearson’s correlation test were evaluated to determine if there was a linear relationship between whale shark sightings and feeding behaviours.

The first three research questions of the study (i, ii, iii) were addressed through a multiple step procedure, named Three-Step System Multivariate Classification (TSMC), across 1082 recorded surface feeding behaviours, 81 independent whale shark sightings, and four time periods (2017, 2020, 2022, and 2024). In this context, a correlation matrix was performed to deal with potential (multi)collinearity problems which might let the predictors be strongly correlated between them (referring to similar events).

The first step used a Permutational Multivariate Analysis of Variance (PERMANOVA) that can (i) evaluate potential heterogeneity among units; (ii) investigate heterogeneous effects among factors and outcomes (surface feeding behaviour); and (iii) deal with potential endogeneity issues. The null hypothesis refers to homogeneity (equivalent dispersion among groups) while the alternative hypothesis refers to heterogeneity (relevant dispersion matters among groups). Estimates were obtained according to the robust Aitchison distance outlined by Martino et al. [39], which made them applicable to all non-negative data including zero, and accounts for both the presence of several categorical variables and the risk of encountering a dummy variable trap. The latter refers to the case in which two or more columns/rows of the matrix containing the predictors are equal between them (linearly dependent vectors), making it impossible to estimate the related regression parameters. The set of covariates was defined as an $\mathrm{n}\ast \mathrm{k}$ matrix $\mathrm{X}$, where $\mathrm{i}=1,\dots ,\mathrm{n}$ and $\mathrm{j}=1,\dots ,\mathrm{k}$ denoted the units and the variables, respectively. There are 9 defined variables: “okta”, denoting light levels; “sea”, referring to sea conditions; “ntemp”, describing SST; “nwspeed”, describing wind speed; “nrain”, denoting rainfall; “ntime”, denoting the time slot scheduled during a sight; “nenso”, referring to the MEI; “nclorop”, denoting the concentration of chlorophyll-a; and “year” referring to the time period. The outcomes of interest, in this step, refer to “dbehaviour”, denoting whale shark surface feeding behaviour evaluated as a dummy variable.

The second step used a multinomial logistic regression to study the relationship between a variable of interest and a set of predictors by assessing the results achieved in PERMANOVA analysis in greater detail. The (multinomial) logistic function described in Equation (1) was:

$${\mathrm{y}}_{\mathrm{i}}=\sum _{\mathrm{j}=1}^{\mathrm{k}}{\mathrm{X}}_{\mathrm{k}\mathrm{i}}{\mathsf{\gamma }}_{\mathrm{k}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(1)

where i = 1, 2, …, n denotes the feeding behaviours (n = 1082), ${\mathrm{X}}_{\mathrm{k}\mathrm{i}}$ refers to the matrix containing all seven predictors evaluated in the first step, ${\mathsf{\gamma }}_{\mathrm{k}}$ represents the regression parameters to be estimated, and ${\mathsf{\epsilon }}_{\mathrm{i}}$ denotes the error term (or causal component).

In the third step, before computing and comparing (sample) marginal effects for every predictor across units over time to investigate the main predictors (or covariates) affecting the outcomes of interest, a discriminant analysis was performed. This analysis focuses on selecting the “best” submodel solution (or combination of predictors) to predict the variable of interest and avoid potential mis-specified variables, due to confounding effects, and potential (multi)collinearity problems among strictly correlated predictors (as are present in this study), where “best” stands for the subset of predictors most closely fitting the data. The estimating procedure takes the name of Best Subset Selection (BSS) analysis and consists of building and, in turn, comparing, several possible regression models based upon an identified set of covariates. The “best” submodel solution corresponds to the one with the lowest Bayesian Information Criterion (BIC). The BSS and the logit regression are classified as Machine Learning (ML) algorithms able to predict output values from a given set of input variables. According to the logit model described in Equation (1), the marginal effects can be calculated in Equation (2) as:

$${\displaystyle \frac{\partial \mathrm{F}\left({\mathrm{X}}_{\mathrm{k}\mathrm{i}}\mathsf{\gamma }\right)}{\partial {\mathrm{x}}_{\mathrm{k}\mathrm{i}}}}=\mathrm{f}\left({\mathrm{X}}_{\mathrm{k}\mathrm{i}}\mathsf{\gamma }\right)\ast {\mathsf{\gamma }}_{\mathrm{k}}$$

(2)

where ${\mathrm{x}}_{\mathrm{k}\mathrm{i}}$ denotes each predictor that is accounted for. In a logit model, marginal effects represent the change in the predicted probability of an event (the choice of the feeding behaviour) due to a small change in one of the explanatory variables (environmental factors), holding other variables constant.

In this context, assuming that there is already a sufficient but not perfect correlation between environmental factors in predicting the feeding behaviour, Odds Ratios (ORs) were computed for each explanatory variable (environmental factors) evaluating the probability of an event favourable to an outcome. They correspond to the exponential of the estimated regression parameters ${\widehat{\mathsf{\gamma }}}_{\mathrm{k}}$, and their related probabilities are computed as $\left(1-\mathrm{O}\mathrm{R}\right)*100$. The utility of defining OR in terms of probability lies in the (i) strong correlation between predictors underestimating the (sample) marginal effects; (ii) multivariate classification based on discrete variables; and (iii) property of the probabilities assuming values from 0 to 1, according to the possibility of reaching infinite values in case of (multi)collinearity problems.

The fourth and last research question was analysed through a confusion matrix, which is a visualisation tool that corresponds to a performance measurement for an ML classification algorithm. In this study, it was applied to multiclass classification problems based on the estimates achieved in the BSS analysis. For instance, the confusion matrix was expanded to a multiclass classification, where each row and column represents the actual classes (environmental factors) and each cell represents the correctly predicted probabilities. One advantage of the confusion matrix is its ability to compute the predictive capability to verify the accuracy (or consistency) of the estimates achieved in a ML classification algorithm. In this section, we consider the variable of interest built as an ordinal variable to understand how sea and weather conditions affected every possible surface feeding behaviour and why whale sharks were inclined to assume active (A), passive (P), or vertical (V) actions. Every confusion matrix was interpreted in terms of probability to better evaluate the results. The elements inside the table denote the joint probabilities. The predictive capability is computed as $1-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}(\%)$, where the mean is obtained by the ratio between the outcomes on the main diagonal (representing the number of successes and failures) and the total observations.

In addition, a Cochran’s Q test was performed to emphasise the strong dependency between whale shark surface feeding behaviour and the other environmental variables. The null hypothesis (${\mathrm{H}}_0$) refers to causality of the surface feeding behaviour, while the alternative hypothesis (${\mathrm{H}}_1$) is applicable where there is an impact on surface feeding behaviour.

3. Results

3.1. Whale Shark Identification and Surface Feeding Behaviour

A total of 81 specimens were photo-identified: 5 in January 2017, 6 in January 2020, 27 in January 2022, 10 in November 2022, and 33 in January 2024. No whale sharks identified were re-sighted between 2017, 2020, and January 2022. In November 2022, 11 whale sharks were re-sighted. Of the re-sightings, 1 was originally sighted in January 2020, and 10 were sighted in January 2022. In 2024, 15 whale sharks were re-sighted. Among these, 2 were from 2020 and 13 were from 2022 (of which 11 were observed in January 2022 and 2 were observed in November 2022). All the individuals observed were males.

A total of 37 whale sharks were measured in size: January 2022 had 10 with an average TL of 6.14 ± 0.1 m; November 2022 had 9 with an average TL of 6.05 ± 0.1 m; and January 2024 had 18 with an average TL of 6.4 ± 0.1 m. Based on the measured TL, these sharks were classified as immature males [40]. Re-sighted and re-measured sharks between 2022 and 2024 (only two specimens) showed an increase of a few centimetres during these two years (14 and 26 cm respectively).

Overall, 1082 surface feeding behaviours were recorded. In January 2017, the most frequent surface feeding behaviour was P (54.91%), followed by V (56.67%) in January 2020, P (44.56%) in January 2022, V (49.49%) in November 2022, and V (53.69%) in January 2024 (

Table 1. Percentages (%) of surface feeding behaviours displayed by whale sharks across sampling years.

Surface Feeding Behaviour	January 2017	January 2020	January 2022	November 2022	January 2024
A	14.01	23.21	19.88	24.62	27.51
P	54.91	20.12	44.56	25.89	18.80
V	31.08	56.67	35.56	49.49	53.69
Total	100	100	100	100	100

(A) stands for active feeding; (P) stands for passive feeding; and (V) stands for vertical feeding.

).

Regarding the chi-square test of independence, where the test statistic with Yates correction was 1.26, with a not statistically significant p-value, we can accept the null hypothesis of independence. This means that there were no differences between surface feeding behaviours affected by environmental factors displayed in January and November 2022.

According to the possible presence of a linear relationship between whale shark sightings and feeding behaviours displayed, the results from the regression analysis and Pearson’s correlation test highlighted that the evidence is not strong enough to confirm a statistically significant dependency between these two variables. Indeed, the estimated regression coefficient is 15.62, with a not statistically significant p-value of 0.307. However, since the R-squared value is 0.48 (hence close to but slightly below 0.50), showing that whale shark sightings explain about 48% of the variance in feeding behaviours displayed, and the Pearson’s correlation coefficient is 0.69, there is evidence of a moderately high positive relationship. This latter could be due to either the small sample size observed for every year or the presence of not directly observed factors and/or whale shark individuality.

3.2. Influence of Environmental Factors on the Surface Feeding Behaviour

Regarding the correlation matrix, variables resulting in a strong correlation with more than one predictor were “nenso” (with “ntemp”, “nwspeed”, and “nrain” displaying a correlation function $\ge 55\%$), and “year” (with “sea”, “nenso”, and “nwspeed” displaying a correlation function $\ge 40\%$).

Thus, the number of variables considered for further analysis was reduced from the original nine to seven potentially significant factors.

3.2.1. PERMANOVA

The first step, the PERMANOVA test (

Table 2. PERMANOVA analysis between environmental factors and whale shark surface feeding behaviour.

Source	df	SS	Pseudo-F	Pr(>F)
okta	1	0.041	7.172	0.00 ***
sea	1	0.025	3.953	0.00 ***
ntemp	1	0.831	5.568	0.00 ***
nwspeed	1	0.193	4.987	0.03 **
nrain	1	1.188	5.525	0.00 ***
ntime	1	0.175	4.991	0.03 **
nclorop	1	0.045	5.015	0.01 **
Residual	1075	0.003
Total	1082	0.005

Here, the labels stand for “degrees of freedom” (df), “Sum of Squared Dissimilarities” (SS), “Pseudo F-statistic” (Pseudo-F), and “associated p-value” (Pr(>F)). The significance levels are: (**) significance at 5%; and (***) significance at 1%.

), shows that the surface feeding behaviour is not random but is affected by potential factor that were not directly observed and measured environmental factors varying over time.

Table 2 displays the estimation outputs, where the significance level is α = 5% (as default). All predictors are significant at least at 5%, displaying a p-value close to zero and lower than α, and the residuals’ standard errors are strictly close to zero, suggesting accuracy and efficiency in the analysis. Looking closely at the estimates, the factor with highly larger Pseudo-F statistics and high significance is “okta” followed by “ntemp” and “nrain”, even if these last two show lower Pseudo-F statistics and larger SS because of similar values collected in the dataset. The value of “nclorop” shows a sufficient significance even if displaying a low Pseudo-F test statistic due to its strong association with sea and weather conditions. For instance, chlorophyll-a concentration is inversely correlated with visibility in terms of “okta” (about 48%) and SST (“ntemp”, about 55%). Also, “nwspeed” and “ntime” show similar results because of their positive correlation (about 55%). Finally, sea conditions positively affect A or V feeding (y = 1), but with a lower Pseudo-F statistic.

3.2.2. Multinomial Logistic Regression Model

The second step, the multinomial logistic regression model (

Table 3. Multinomial logistic regression functions analysing how environmental factors affect whale shark surface feeding behaviour across units over time.

Coefficients	Estimate	SE	z-value	Pr(>|z|)
okta	3.051	0.038	80.29	0.00 ***
sea	3.033	0.128	23.70	0.00 ***
ntemp	−2.023	0.135	−14.99	0.00 ***
nwspeed	−1.302	0.141	−9.23	0.00 ***
nrain	3.438	0.145	23.71	0.00 ***
ntime	−1.652	0.141	−11.72	0.00 ***
nclorop	2.405	0.201	11.97	0.00 ***

Here, “Coefficients” refers to the covariates; “Estimate” refers to ${\widehat{\mathsf{\gamma }}}_{\mathrm{k}}$ (the estimated regression parameters ${\mathsf{\gamma }}_{\mathrm{k}}$); “SE” stands for Standard Error; “z-value” denotes the test statistic obtained for each predictor (the ratio between “Estimate” and “SE”); and “Pr(>|z|)” refers to the associated p-value according to a two-sided hypothesis testing (where the null stands for non-significance). The significance levels are: (***) significance at 1%.

), highlights how the main environmental factors drive whale sharks to choose the filter-feeding strategy.

Table 3 displays the estimation outputs, where the significance level is α = 5% (as default). All predictors are significant at 1%, highlighting the efficiency of the supervised ML algorithm at minimising the sum of squared residuals and avoiding potential (multi)collinearity. Indeed, the full (or unrestricted) model, with all nine predictors, shows an Akaike Information Criterion (AIC) equal to 333.40, while the restricted one (without “nenso” and “year”) is equal to 311.89. The lowest AIC is the most preferred (restricted model).

Looking at the Table 3, the values “okta”, “sea”, “nrain”, and “nclorop” have a positive z-value, while “ntemp”, “nwspeed”, and “ntime” have a negative z-value, indicating a significant effect of predictors on whale shark surface feeding behaviour. Higher values of “sea” correspond to worse sea conditions. The predictor “nrain” should be interpreted with care because the maximum mm/h collected during the sightings was 0.1, where less rainfall was observed with A or V feeding behaviours. Factors such as lower light levels (cloudier sky), rougher sea, rainfall, and higher chlorophyll-a levels positively affect the whale shark feeding behaviour towards A or V strategies (y = 1). On the contrary, higher SST, strong wind speed, and afternoon hours negatively affect the whale shark feeding behaviour towards the P strategy (y = 0).

3.2.3. Best Subset Selection (BSS)

A BSS (Figure 2) was performed to better evaluate (and then confirm) the results obtained in the multinomial logistic regression.

According to Figure 2, the “best” subset of predictors corresponds to the ones with lower BIC (positive values). Therefore, the best subsets inclining whale sharks to assume A or V feeding behaviours are associated with values of SST $\le 26.2$ °C, concentration of chlorophyll-a $\ge$ 0.60 mg/m³ (assuming values 2 and 3 by construction), rainfall $\le$ 0.1, and ENSO measurement unit $\le$−2.1 (good weather conditions). Chlorophyll-a with a value 0.473 mg/m³, even if displaying more black squares than the values $\ge$0.60 mg/m³, was discarded from the discriminant analysis because of its significance at a highly larger BIC as well (from 23 to 51). These findings confirm and deepen the estimation results found in Table 3.

3.2.4. Odds Ratios and Sample Marginal Effects

The third step, the sample marginal effects (

Table 4. Sample marginal effects for each observation unit, given n observations, are accounted for.

Coefficients	dF/dx	SE	z-value	Pr(>|z|)
okta	0.430	0.143	3.01	0.00 ***
sea	0.286	0.127	2.25	0.01 **
ntemp	−0.415	0.172	−2.41	0.01 **
nwspeed	−0.276	0.131	−2.11	0.02 **
nrain	0.367	0.176	2.09	0.02 **
ntime	−0.201	0.115	−1.75	0.04 **
nclorop	0.413	0.141	2.93	0.00 ***

Here, “Coefficients” refers to the factors within the model; “dF/dx” denotes the partial derivatives displaying the marginal effects of the predictors (${\mathrm{x}}_{\mathrm{i}\mathrm{k}}^{\prime }$) on ${\mathrm{y}}_{\mathrm{i}}$ (“dbehaviour”); “SE” stands for Standard Error; “z-value” denotes the test statistic obtained for each predictor; and “Pr(>|z|)” refers to the associated p-value in a two-sided hypothesis test (where the null accounts for non-significance). The significance levels are: (**) significance at 5%; and (***) significance at 1%.

), evaluates major environmental factors affecting the surface feeding behaviour.

Table 4 displays the estimation outputs, where the significance level is α = 5% (as default). All predictors are significant at least at 5%, According to the estimates displayed in Table 4, the main factors affecting whale shark surface feeding behaviour, in order of importance, are “nclorop” (79.80%), “okta” (64.15%), “nrain” (62.50%), “ntemp” (59.26%), “sea” (49.66%), “nwspeed” (44.24%), and “ntime” (26.59%). The OR results fully correspond to the ones found in PERMANOVA analysis. Finally, a White’s heteroskedasticity correction test was performed to standardise the residuals dealing with potential (multi)collinearity problems.

3.2.5. Confusion Matrices

A confusion matrix was then used for each of the main factors affecting whale shark surface feeding behaviour: “nclorop”, “okta”, “rain”, and “temp”. The last three predictors (“sea”, “nwspeed”, and “ntime”) were evaluated through ENSO measurement unit and discarded from the analysis because of their strong correlation. In this context, SST and the MEI were not considered grouped in classes to investigate their possible value.

The first confusion matrix (

Table 5. Confusion matrix between surface feeding behaviour and chlorophyll-a.

Behaviour/Chlorophyll-a	1	2	3
1 (A)	0.187	0.614	0.199
2 (P)	0.703	0.123	0.174
3 (V)	0.223	0.637	0.140

The values inside the table correspond to the joint probabilities between two discrete variables (Y and X) for each possible outcome. Here, Y refers to whale shark feeding behaviour (A stands for active feeding; P for passive feeding, and V for vertical feeding) and X refers to the number of chlorophyll-a grouped in classes.

) highlights how chlorophyll-a concentration affects the surface feeding behaviour.

Results displayed in Table 5 show that either A or V feeding behaviour has the highest probabilities when chlorophyll-a concentration assumes the value of 2 (between 0.51–2.00 mg/m³ by construction). Conversely, the P strategy has the highest probability when chlorophyll-a concentration assumes a value of 1 or 3. Thus, in the presence of massive chlorophyll-a concentrations (assigned a value of 3), whale sharks would tend to change their feeding behaviour from A/V to P, but only if associated with favourable sea and weather conditions (according to PERMANOVA results). The predictive capability is 83.20%.

The second confusion matrix (

Table 6. Confusion matrix between surface feeding behaviour and okta.

Behaviour/Okta	0	1	2	3	4	5	6	7	8
1 (A)	0.014	0.017	0.021	0.032	0.439	0.055	0.073	0.106	0.261
2 (P)	0.102	0.120	0.101	0.068	0.505	0.035	0.021	0.062	0.107
3 (V)	0.010	0.012	0.048	0.024	0.445	0.050	0.062	0.136	0.225

The values inside the table correspond to the joint probabilities between two discrete variables (Y and X) for each possible outcome. Here, Y refers to whale shark feeding behaviour (A stands for active feeding; P for passive feeding, and V for vertical feeding) and X refers to light levels measured through the predictor “okta”.

) highlights how light levels affect the surface feeding behaviour.

Results displayed in Table 6 show that when oktas increased up to the maximum value (=8), A and V feeding behaviours are favoured, exceeding the P feeding strategy. Conversely, better light levels (lower oktas) correspond to P feeding behaviour. The predictive capability is 84.46%.

The third confusion matrix (

Table 7. Confusion matrix between surface feeding behaviour and rainfall.

Behaviour/Rainfall	0	1
1 (A)	0.005	0.996
2 (P)	0.069	0.931
3 (V)	0.035	0.965

The values inside the table correspond to the joint probabilities between two discrete variables (Y and X) for each possible outcome. Here, Y refers to whale shark feeding behaviour (A stands for active feeding; P for passive feeding, and V for vertical feeding) and X refers to rainfall.

) highlights how rainfall affects the surface feeding behaviour.

Results displayed in Table 7 show that the total absence of rain (=0) is associated with the P strategy, while more rainfall (=1) corresponds to A and V strategies. The predictive capability is 88.80%.

The fourth confusion matrix (

Table 8. Confusion matrix between surface feeding behaviour and SST.

Behaviour/SST	26.1	26.2	28.07
1 (A)	0.012	0.894	0.094
2 (P)	0.058	0.069	0.873
3 (V)	0.039	0.143	0.818

The values inside the table correspond to the joint probabilities between two discrete variables (Y and X) for each possible outcome. Here, Y refers to whale shark feeding behaviour (A stands for active feeding; P for passive feeding, and V for vertical feeding) and X refers to SST.

) highlights how SST affects the surface feeding behaviour.

Results displayed in Table 8 show that the optimal SST value is >26 °C. Thus, the SST values used are 26.1 °C and 26.2 °C, as found in BSS analysis, and 28.07 °C as the mean in 2024 (where the highest SST was recorded). An increase in SST ($>$26.2 °C) would favour P and more V than the A feeding behaviour. For instance, these findings highlight that V and P feeding behaviours are more favoured when the SST is sufficiently high, particularly with values being higher than average SST (26.2 °C). Conversely, the A feeding behaviour is favoured with an average SST of 26.2 °C. The predictive capability is 88%.

The last confusion matrix (

Table 9. Confusion matrix between surface feeding behaviour and the MEI.

Behaviour/MEI	−2.1	−0.3	2.0
1 (A)	0.811	0.075	0.116
2 (P)	0.072	0.009	0.920
3 (V)	0.053	0.029	0.926

The values inside the table correspond to the joint probabilities between two discrete variables (Y and X) for each possible outcome. Here, Y refers to whale shark feeding behaviour (A stands for active feeding; P for passive feeding, and V for vertical feeding) and X refers to the ENSO measurement unit expressed through the MEI.

) highlights how ENSO affects the surface feeding behaviour.

Results displayed in Table 9 show that values of the MEI $\le$ −2.1 predict an A feeding strategy (just as found in BSS analysis). Conversely, an increase in the MEI is associated with V and, to a lesser extent, P feeding behaviours. The predictive capability is 88.4%.

The Cochran’s Q test statistic is −211.80, supporting the dependence between variables and rejecting the null hypothesis (p-value close to zero). The significance level is 5% as default, with one degree of freedom.

4. Discussion

During the sampling period, 81 different whale sharks were identified through the I³S Classic program, which confirmed that both whale shark flanks above the pectoral fins, just posterior to the gill slits, had patterns unique to each shark that remained unchanged over time [36,37,41]. This finding further confirms that Djibouti hosts an important population of whale sharks, as indicated by Boldrocchi et al. [6], who photo-identified 190 whale sharks between 2015 and 2018, and by Rowat et al. [3], who reported 297 individuals identified between 2003 and 2010.

In addition, 26 whale sharks (32.10% of the whale sharks identified) were re-sighted across years supporting the conclusion by Boldrocchi et al. [6] that some individuals had a prolonged residency in this region. Although the sites where whale sharks move during the off-season from March to September are currently unknown, satellite tag data from Rowat et al. [42] showed that two whale sharks travelled from Djibouti into the Red Sea and the northern Indian Ocean, and the reasons for these two migrations remain unknown. However, data were too limited to prove a migration to other aggregation sites, suggesting that their dispersion from Djibouti was minimal [6].

Our results suggested an increase in sightings over time, going from a few whale sharks identified in January 2017 (5 specimens) and January 2020 (6 specimens) to a higher number in 2022 (27 specimens in January and 10 in November) and January 2024 (33 specimens). It is known that whale shark abundance and distribution are associated with food availability at coastal locations [3,6,17,43], where local environmental conditions support the species and zooplankton biomass [28]; for instance, the zooplankton biomass in the Gulf of Tadjoura increased from 24.8 ± 9.1 mg/m³ [11] in 2017 to 42.2 ± 31.9 mg/m³ in 2018 [6]. While the annual concentration of zooplankton biomass was not provided in this study, a positive relationship between monthly chlorophyll-a concentration and zooplankton biomass collected during the whale shark aggregation period was observed by Boldrocchi et al. [6] in the Gulf of Tadjoura. Thus, Copernicus satellite data of chlorophyll-a concentration for the study area were used. The collected data showed a constant increase in chlorophyll-a concentration in 2020 (0.47 mg/m³), 2022 (0.53 mg/m³), and 2024 (0.61 mg/m³) compared to 2017 (0.25 mg/m³), which may have enhanced the zooplankton abundance and, in turn, facilitated an increase in the number of whale shark sightings over time. Although whale shark sightings seem to not affect feeding behaviours, the moderately high positive relationship between these two variables suggests that a bigger sample size, whale shark individuality, and other not directly observed factors must be considered in future sampling years.

All whale sharks measured in TL were immature males. Even the remaining whale sharks whose sizes were not assessed showed visually very similar TLs; thus, they were considered immature males as well. These findings confirm previous reports by Rowat et al. [3] and Boldrocchi et al. [6] that there is a male-based aggregation in Djibouti, which is the phenomenon also common in other Indian Ocean aggregation sites [3,44,45]. However, although sex- and size-based segregation are common among shark species due to different nutritional requirements or to avoid intra-specific competition and predation, the reason for sexual segregation amongst whale sharks remains unclear; different diet preferences between immature and mature whale sharks have been suggested [45].

Environmental factors significantly influenced the different surface feeding behaviours exhibited by immature male whale sharks over different years.

Satellite-measured chlorophyll-a was the most important environmental factor predicting the choice of the surface feeding behaviour for immature male whale sharks in the study areas, but only if considered in direct association with the other sea and weather conditions. Chlorophyll-a concentration is known to be a proxy of phytoplankton biomass, which in turn affects zooplankton availability [46]. Several studies have demonstrated the positive correlation between chlorophyll-a concentration and zooplankton abundance [6,28,29,30]. Thus, the chlorophyll-a concentration could be the predicting factor in the choice of the surface feeding behaviour for immature male whale sharks, although a detailed study on the quantities of chlorophyll-a and the corresponding quantities and diversity of daily plankton in Djiboutian waters is necessary to better understand a possible interaction effect on feeding behaviour. In addition, it is known that when whale sharks are identifying appropriate areas in which to feed, they may exhibit a foraging response to the volatile chemical dimethyl sulphide (DMS), which is released when zooplankton graze on phytoplankton [47]. Thus, it will be also useful to understand if surface feeding behaviour is affected not only by chlorophyll-a concentration related to zooplankton biomass, but also by different phytoplanktonic families that release the DMS when zooplankton graze on them.

In bad sea and weather conditions, chlorophyll-a concentration $\ge$0.60 mg/m³ could drive whale sharks to exhibit A or V feeding behaviours in the presence of expected high or medium zooplankton biomass. On the contrary, in good sea and weather conditions, with lower (≤0.60 mg/m³) or higher concentrations of chlorophyll-a ($\ge$2.01 mg/m³), whale sharks exhibit a shift in their feeding behaviour from A or V to the P feeding strategy because it requires less effort on the part of the whale sharks [48]. Indeed, the P feeding strategy is a means to subsidise energy efforts while the whale shark is searching for a higher density of food. In Djibouti, Di Capua et al. [11] also stated that the P feeding strategy occurred with the lowest abundance of zooplankton, while the A and V feeding behaviours were associated with a higher abundance. Nelson and Eckert [22] observed a similar trend in Bahìa de Los Angeles (Mexico), where whale sharks exhibited an A feeding strategy with high, V with medium, and P with low zooplankton biomass. Similarly, basking sharks (Cetorhinus maximus, Gunnerus, (1765)) have been shown to adjust their feeding strategies in response to environmental predictors such as chlorophyll-a concentration and abundance of prey [49]. For example, they actively forage for plankton to make the effort worthwhile when prey densities are above a particular value, showing them to be a selective filter-feeder [50].

In this context, given that chlorophyll-a concentration was the most important environmental factor influencing the feeding choice, it is important to understand how its concentration is regulated and affected by the other environmental factors.

Worse climatic conditions are associated with whale sharks feeding in A or V, mostly because the concentration of chlorophyll-a tends to increase with higher rainfall and cloudier conditions (higher oktas values) related to a worse underwater visibility. Indeed, light levels having higher oktas values positively affect the probability of A and V feeding behaviours. It means that with cloudier skies, immature male whale sharks perform more A and V feeding behaviours. Furthermore, rainfall also positively affects the probability of A and V feeding behaviours, and immature male whale sharks perform more A and V feeding behaviours with higher rainfall. Since immature male whale sharks perform more A and V feeding behaviours with higher chlorophyll-a concentration under worse climatic conditions, the probability of the occurrence of these feeding behaviours and also of a higher chlorophyll-a concentration increases with cloudier skies and greater rainfall. Hacohen-Domené et al. [28] suggested that chlorophyll-a was positively correlated with rainfall, which increases the primary productivity and the presence of whale sharks in Cabo Catoche and Isla Contoy (Mexico). Given that the whale shark is known to rely on hearing and olfactory senses rather than visual acuity when locating and capturing prey [47], light levels do not play a significant role in visually locating the prey. However, lower light levels (higher oktas values) usually correlate with a higher rainfall rate and, consequently, increased chlorophyll-a events.

Regarding sea conditions, chlorophyll-a was inversely correlated with SST, suggesting that as the SST increases, chlorophyll-a concentrations decrease. The upwelling phenomenon is one of the most important oceanographic conditions for increasing food availability [28,30,51,52]. The upward movement of cold waters increases nutrients, chlorophyll-a concentrations and, consequently, zooplankton biomass [53]. During autumn, the northeast monsoon generates an increase in chlorophyll-a that coincides with colder SST and induces an upwelling event along the southern coastline of the gulf of Somalia, causing a nutrient enrichment [54]. Upwelled cold waters converge with the warm waters of the Red Sea, creating ideal conditions for the proliferation of phytoplankton and, consequently, zooplankton biomass along the coastline of Djibouti. As a result, whale sharks in Djibouti find rich feeding grounds, actively filtering water to consume vast quantities of prey. However, as SST rises, whale sharks switch to V or P feeding behaviours, which are more energy-efficient in these environmental conditions. In fact, predicted rises in SST would increase whale shark metabolic demands, potentially affecting its energy requirements and feeding behaviour as available prey change [23]. Indeed, Hacohen-Domenè et al. [28] found that presence of whale sharks in Cabo Catoche and Isla Contoy (Mexico) may be driven by an increase in temperature triggering the spawning of little tuna [43], given that these areas are characterised by lower primary and secondary production. However, this aspect was not investigated in this study, and further analyses will be necessary to assess if the increase in SST affects immature male whale shark feeding behaviour in Djibouti as the availability of prey changes. To date, this phenomenon does not seem to be the case for immature male whale sharks in Djibouti, since they feed mainly on high concentrations of zooplanktonic prey [6,10,11]. However, only seven whale sharks in Djibouti have been observed feeding on a school of baitfish (anchovies) on 16th October 2017 and during the off-season, when dense patches of zooplankton were not available [13]. Thus, further studies are required to assess whether immature male whale sharks in Djibouti also feed on schooling fish during the high season of the feeding aggregation and if this phenomenon could be linked to the increase in SST and, consequently, the change in prey availability.

The wind speed was also inversely correlated to chlorophyll-a concentration, meaning that stronger winds would disperse and, consequently, decrease the chlorophyll-a concentration at the surface during the study period, and immature male whale sharks tended to use the P feeding strategy. Indeed, wind speed negatively affects the probability of A and V feeding behaviours. It means that with stronger wind speed, immature male whale sharks perform more P feeding behaviour. Conversely, chlorophyll-a concentration positively affects the probability of A and V feeding behaviours and it means that with higher chlorophyll-a concentration, immature male whale sharks perform more A and V behaviours. Thus, stronger wind speed and lower chlorophyll-a concentration matter with P feeding behaviour under unfavourable weather conditions. Negative effects of wind speed on the number of whale sharks recorded in Seychelles have been further reported by Rowat et al. [55], showing that higher wind speed causes surface disturbance and prompts a change in shark behaviour. However, from aerial surveys, the authors were not able to determine if sharks had left the area or went to lower depths beyond sighting range. It is possible that in Djibouti, stronger winds disperse dense zooplankton patches, inducing immature male whale sharks feeding with A or V strategies to leave these behaviours and change to the P strategy just below the surface. Since worse sea conditions are a consequence of stronger wind speeds, whale sharks may adopt a P feeding strategy in rough seas. However, worse sea conditions increased the concentration of chlorophyll-a, and slightly rough seas drove sharks to use A or V feeding behaviours. Since both sea condition and wind speed were of less importance in influencing the surface feeding behaviour, it is possible that whale sharks adopted a P feeding strategy when the sea was calm and the weather was good with high density patches of zooplankton to reduce their energy efforts and increase energy gain from prey. Conversely, more expensive A or V feeding strategies were preferred with slightly rough seas and moderate wind speed, suggesting that whale sharks forage more aggressively in these conditions.

The time of the day was the last environmental factor to influence the choice of the filter-feeding strategy. During the morning with worse sea conditions, lower SST, higher rainfall, higher concentration of chlorophyll-a, and low wind speed, immature male whale sharks tended to exhibit A or V feeding behaviours at the surface, probably because dense patches of prey were more available during this time, while during the afternoon they preferred the P strategy. Since diel patterns in surface feeding activity in Djibouti have never been described, it is difficult to assess if surface feeding behaviours follow the horizontal and/or vertical diel changes in prey distribution in this area. Decreased light levels during the afternoon (14.00–17.00) with respect to the morning (09.00–12.00) are thought to reduce predation pressure on zooplankton by visual predators [31,56]. Furthermore, Motta et al. [19] observed a peak in abundance of whale sharks filter-feeding during mid-morning in Cabo Catoche (Mexico); on the contrary, Gleiss et al. [31] stated that whale sharks at Ningaloo Reef exhibited ram filter-feeding techniques primarily during sunset and the first hours of the night. Thus, pronounced phases of filter-feeding techniques suggest that temporal dynamics of zooplankton aggregation are critical factors in influencing the behaviour of whale sharks in Djibouti.

Besides sea and weather conditions influencing filter-feeding strategies, the role of the ENSO phenomenon on the surface feeding behaviour must also be taken into consideration. In Djibouti, whale sharks utilised more A feeding strategies during La Niña and more P or V during El Niño events. ENSO is a global ocean–atmosphere process causing global climate variability on seasonal to interannual time scales [57] and it acts as an indirect proxy for wind and rainfall [58]. Warm events are characterised by El Niño, while cold ones are characterised by La Niña [57,59]. ENSO is an important component of climate variability to take in consideration along the Djiboutian coast, which is known for its arid to semi-arid climates and extreme interannual rainfall variations [60]. El Niño events seem to correspond to wet years in Djibouti [61], driving a northward shift of the Red Sea Convergence Zone in October–January, with more frequent and more intense rainy days over the Red Sea and adjoining regions associated with strengthened easterlies and southeasterlies winds over the Gulf of Aden and southern Red Sea, warmer SSTs, and higher convective instability [62].

However, V and P were the most frequent feeding strategies observed during the strong El Niño event in 2024 (MEI 2) and the weaker one in 2017 (MEI 0.5), as well as during La Niña events occurring in 2020 (MEI −0.2) and 2022 (MEI −0.9). Feeding strategy A has never been recorded as the most frequent behaviour during La Niña events. Although ENSO and related environmental variables (SST, rainfall, and wind speed) play an important role in influencing the abundance of chlorophyll-a and, consequently, the surface feeding behaviour, it seems that other factors, such as currents and vertical prey abundance variations, must be taken in consideration to better predict the choice of the feeding behaviour. However, this environmental parameter was not considered in this study. Other authors highlighted that, during La Niña, trade winds drive stronger currents and warmer sea temperatures along the north coast of Australia, positively influencing the southward flow of the Leeuwin Current along the west coast of Australia [26,32]. Thus, the Leeuwin Current is likely to be very important in producing localised re-suspension of nutrients and productivity pulses [26]. Manuhutu et al. [30] also suggested that whale shark abundance in Cenderawasih Bay is influenced by the Indonesian Throughflow, which transports water rich in nutrients from the Pacific to the Indian Ocean associated with ENSO events. Furthermore, abundances and distributions of basking sharks have also been linked with the North Atlantic Oscillation off the coast of southern England [50] and ENSO events off the coast of California [63]. The same phenomenon could also happen along Djibouti coast, where specific currents driven by ENSO events could influence the concentration of chlorophyll-a in the area and, consequently, both the abundance of immature male whale sharks and their choice of the feeding behaviour.

5. Conclusions

Studying the effects of environmental factors on the whale shark surface feeding behaviour is challenging, especially given the lack of information in the literature concerning the topic. Findings suggest that chlorophyll-a concentration, influenced by other environmental factors, is a key driver of immature male whale shark feeding behaviour, although several aspects related to the consequential phyto- and zooplankton concentrations must be improved. Since the whale shark is particularly vulnerable to changes in food availability, global climate change could have a significant impact on its feeding ecology as well as its distribution across different hot spots of the world. Therefore, it is recommended that future analyses with longer time series should be performed to investigate other environmental parameters that could influence the choice of the filter-feeding technique and which were not considered in this study. Predicting possible feeding behaviours for this species through environmental factors can be used to evaluate the possible habitat distribution of whale sharks in this area, and understanding habitat utilisation is essential for management decisions.

This research is expected to contribute to the lack of knowledge on the feeding behaviour of immature male whale sharks found in Djibouti during this study and become an initial reference in managing this species within this area.