Overharvesting, bycatch and pollution were the main anthropogenic impacts on cetaceans and sea turtles worldwide during the second half of the 20^th century and pushed some species and populations to the brink of extinction (Wallace et al., 2011; Mittermeier and Wilson, 2014). As new regulations were enforced, the impact of such threats decreased and many populations of both groups began to recover during the first decades of the 21^st century (Duarte et al., 2020). However, the unprecedented increase of maritime traffic (Walker et al., 2019) and the development of new technologies such as offshore wind farms and deep-sea mining represent new challenges for their conservation due to the potential loss of suitable habitats (Maxwell et al., 2022).

The Mediterranean Sea supports intense human activity and has experienced a major erosion of its biodiversity (Coll et al., 2012). However, it still supports 13 breeding species of marine mammals and two nesting species of sea turtles (Notarbartolo di Sciara, 2016; Casale et al., 2018). The distribution of cetaceans and sea turtles within the basin is highly heterogeneous and the highest at-sea species diversity has been recorded in the western Mediterranean Sea (DiMatteo et al., 2022; Cañadas et al., 2023). This is because of the higher primary productivity of the western basin compared to the eastern Mediterranean Sea (Bosc et al., 2004), as food availability for air-breathing marine vertebrates is determined mostly by primary productivity. Nevertheless, the major nesting beaches of sea turtles are found in the eastern Mediterranean Sea, because of the higher sand temperature during the summer months (Casale et al., 2018).

Studies conducted during the 1990s and early 2000s identified the Balearic Sea (between the Iberian Peninsula and the Balearic Islands), as a hotspot for the conservation of cetaceans (Forcada et al., 1994, Forcada et al., 1996, Forcada et al., 2004; Raga and Pantoja, 2004; Gómez de Segura et al., 2006a and Gómez de Segura et al., 2006b). Particularly compelling was the evidence that the spring migration of fin whales, connecting their still unknown breeding grounds, somewhere in the southern part of the Mediterranean basin, and their foraging grounds in the Gulf of Lions and the Ligurian Sea, run parallel to the Iberian Peninsula’s coastline (Raga and Pantoja, 2004). As a result, the Spanish Government designated in 2018 the Mediterranean Cetacean Migration Corridor as a Marine Protected Area and proposed to the Barcelona Convention the inclusion of the area in the SPAMI List (Specially Protected Area of Mediterranean Importance). The extent of the area is 46,385.70 km² and encloses roughly the areas in between the shelf break of the Iberian Peninsula and the shelf break west to the Balearic Islands, from latitude 38° 39’ 59.379’’ to latitude 42° 18’ 57.294’’ ( Figure 1 ). Oil extraction and mining have been forbidden within the area, but the limited information on the distribution and abundance of cetaceans and marine turtles in this Marine Protected Area has precluded the development of specific regulations on maritime traffic, although collisions have been identified as a potential threat for cetaceans in the area (Panigada et al., 2006 and Panigada et al., 2017b; Tort Castro et al., 2022).

Population estimates and species distribution models are critical pieces of information for the management of cetaceans and sea turtles, with aerial and boat surveys being the most common method used to gather this information. Population estimates are usually conducted within the framework of distance sampling (Buckland et al., 1993; Thomas et al, 2010), whereas species distribution models (SDMs) can be developed using different approaches. The most basic distinction is between presence-absence and presence-only models. The choice of the model depends on the type of data available, particularly whether data on survey effort was recorded alongside species sightings (Guisan and Zimmermann, 2000). Presence-absence models require both presence and absence data collected during planned surveys. Each confirmed sighting within the survey area represents a successful detection of the target species under investigation. Presence-absence models encompass a variety of statistical techniques, including Generalized Linear Models (GLMs), Generalized Additive Models (GAMs), occupancy models, or machine learning algorithms such as boosted regression trees, random forest, or support vector machines (Guisan and Zimmermann, 2000; MacKenzie et al., 2002; Brotons et al., 2004; Reisinger et al., 2021). It should be noted that combining several approaches is strongly recommended to better understand the drivers of species distribution, as modeling approaches may differ in their capacity to capture environmental gradients and the way they select the appropriate variables (Fukuda et al., 2013; Kosicki, 2020).

Random Forest is a method particularly suited to address ecological modeling problems, although it is not inherently a presence-absence model. It is a general machine learning algorithm that can be used for a variety of tasks, including classification, regression, and unsupervised learning. In the context of species distribution modeling, random forest can be used to model both presence-absence and presence-only data. Random forest models (RF) have been increasingly applied in ecological studies due to their robustness and ability to handle complex interactions among predictor variables. For cetaceans, RF have shown promise in accurately predicting species distributions and identifying important environmental variables. For example, Oppel et al. (2012) used RF to model cetacean distributions in the Mediterranean Sea, demonstrating its effectiveness in handling large and complex datasets. Similarly, Scales et al. (2017) employed RF models to predict the distribution of baleen whales in the Northeast Atlantic, highlighting the utility of machine learning approaches in marine conservation.

This study aims to improve our understanding on the abundance and distribution of cetaceans and sea turtles within the Mediterranean Cetacean Migration Corridor by developing species distribution models (SDMs) to identify areas of suitable habitat. By utilizing a combination of presence-absence models, such as GAMs and RFs, we can leverage the strengths of each approach to provide a comprehensive picture of species distribution. This information will serve as a baseline for the adaptive management of human activities in the area, ensuring the conservation of these key marine species while balancing the sustainable use of marine resources.

The study area is located in the northern part of the Mediterranean Cetacean Migration Corridor and encompasses 32,674 km², i.e. 70% of the MPA ( Figure 1 ).

The abundance of cetaceans and sea turtles was assessed through the combination of an aerial survey and a boat survey using Distance Sampling methodology (Thomas et al, 2010). This approach uses the perpendicular distances of individuals or groups of observed animals from the observation platform to calculate a detection function, used later to model the probability of detecting an animal given its distance from the transect line. The effective bandwidth on each side of the transect (i.e., the distance within which animals can be detected) is estimated using this detection function. Once the detection function is calculated, the estimated density can be extrapolated to the entire study area​. Linear transects across the study area were designed using the software DISTANCE 7.3 (Buckland et al., 1993) in order to obtain the same likelihood of coverage within the entire area.

The total number of sightings during the surveys is shown in Table 5 and their distribution in Figure 2 . A total of 113 sightings were recorded during the aerial survey and a total of 268 sightings during the boat survey. Hard-shelled turtles (n = 124) and striped dolphins (n = 225) were the most often spotted species during both surveys, followed by fin whales (n = 44). Risso’s dolphins (n = 19), sperm whales (n = 20), Cuvier’s beaked whales (n = 16) and bottlenose dolphins (n = 5) were also spotted during both surveys, in lower numbers. Long-finned pilot whales were spotted only once during the boat survey. Risso’s dolphins were spotted more often during the aerial survey than the boat survey; the opposite was true for sperm whales and Cuvier’s beaked whales.

The results reported here revealed that the community of air-breathing marine predators inhabiting the northern part of the Mediterranean Cetacean Migration Corridor was dominated numerically by species with a small body mass (striped dolphins and hard-shelled turtles), whereas two large-body species, the fin whale and the sperm whale, concentrated most of the biomass in the community. Nevertheless, the species considered differed largely in their habitat suitability patterns and hence the composition of the community was expected to change within the study area. Finally, the performances of RF and GAM varied largely across the studied species, without any obvious relationship with sample size and with particularly low performances of RF for fin whales and Cuvier's beaked whales. This variability probably emerged because the species differed in their responses to environmental gradients and because RF and GAM differed in the way they selected the environmental predictors to be included and their ability to capture the relevant environmental gradients at the spatial resolution considered. This is because, after accounting for collinearity, RF selected the environmental predictors that maximized the performance of the model, whereas GAM included only the environmental predictors with a statistically significant effect. As a result, RF and GAM did not always select the same environmental predictors. Furthermore, the number of sightings was small, even for the two most abundant species, which is the likely reason for the modest AUC values of both RF and GAM ( Table 8 ) and the low values of deviance explained by GAM. Nevertheless, the performance of the resulting species distribution models were largely similar, except for fin whales and Cuvier's beaked whales, and agreed in identifying the same suitable areas for every species, except for fin whales.

It should be noted that balancing the presences and absences was necessary to calculate AUC and hence enable direct numerical comparison between the performance of the RF and GAM species distribution models. This is because absences were much more abundant than presences in the original dataset and such imbalance might result into very high AUC values if the models successfully predicted absences but failed to predict presences. This is particularly relevant for the binomial GAM used in the present study, as they are based on a probabilistic framework that models the probability of the response variable (binary) using a logit link function. The performance of these models is typically assessed using metrics such as the percentage of deviance explained, which quantifies how much of the variation in the response is accounted for by the model, relative to a null model. This metric is particularly informative when the goal is to interpret the effect of individual predictors within a statistically structured model. However, this metric is unsuitable to assess the performance of RF models, as the latter are non-parametric machine learning models and no particular data distribution is assumed. Hence, AUC is the standard performance metric for this kind of classification models.

In the light of AUC, the performance of RF and GAM was rather similar, except for fin whales and Cuvier's beaked whales. Nevertheless, the final goal of using two modelling approaches was not comparing their performances, but to improve our understanding of the drivers of species distribution and make more robust estimates of habitat suitability (Fukuda et al., 2013; Kosicki, 2020). Here, by combining random forest and generalized additive models we highlighted that the habitat suitability of the species with a small body mass (hard-shelled turtles and striped dolphins) was tightly linked to parameters related to the abundance of primary producers (the concentration of chlorophyll-a or the net primary productivity), whereas that of deep divers (Risso’s dolphins, sperm whales and Cuvier’s beaked whales) was best explained by a combination of bathymetric (standard deviation of the slope, depth or hillshade) and hydrographic (eddy kinetic energy, sea surface height and eastward or northward sea water velocity) variables. Furthermore, the distribution of fin whales was poorly predicted by the environmental variables considered, likely because they were migrating along a well-defined geographic area, independently on the environmental conditions. In any case, during the spring of 2023, most of the habitats suitable for marine mammals and hard-shelled turtles within the study area concentrated along the slope of the Iberian Peninsula and the Balearic Islands, with a remarkable low habitat suitability for the oceanic expanse approximately midway the Ebro Delta and Menorca Island.

It is important noting that the spring aerial survey was the basis of most of the species distribution models presented here. This is particularly relevant for fin whales and loggerhead turtles, known to migrate seasonally (Casale et al., 2018; Panigada et al., 2017a and Panigada et al., 2024). Conversely, no seasonal migration has been described for the remaining species and hence the spring model is probably a good proxy of the year-round distribution. On the other hand, the models for Cuvier’s beaked whales and sperm whales combined data collected during the spring and late summer, for three reasons: only a few sightings were made during the aerial surveys, the spring boat cruise was restricted to the western strata due to bad weather in the eastern strata and the number of sightings during the spring boat cruise was small. Combining observations three months apart might be meaningless for migratory species, but there is no evidence that Cuvier’s beaked whales and sperm whales migrate seasonally and hence there is no reason not to increase sample size by combining observations from spring and late summer.

Previous studies reported a positive relationship between the concentration of chlorophyll-a and the population density of striped dolphins, hard-shelled turtles and fin whales at several foraging grounds in the Mediterranean Sea (Laran and Gannier, 2008; Panigada et al., 2008; Druon et al., 2012; DiMatteo et al., 2022; Tort Castro et al., 2022; Panigada et al., 2024). Hard-shelled turtles and fin whales are epipelagic predators (Panigada et al., 1999; Alvarez de Quevedo et al., 2013; Friedlaender et al., 2020) and this is also probably true for striped dolphins, in the light of the behavior of the phylogenetically close pantropical spotted dolphin (Stenella attenuata; Scott and Chivers, 2009). Oceanic loggerhead turtles rely heavily on gelatinous zooplankton (Revelles et al., 2007b; Cardona et al., 2012), whereas striped dolphins consume mostly mesopelagic fishes and squid (Aznar et al., 2017) and fin-whales primarily feed on krill (Friedlaender et al., 2020). Interestingly, most of the biomass of those forage species concentrates at the deep scattering layer during the light hours, becoming accessible to epipelagic predators only at night, when some zooplankton and mesopelagic fish species migrate to the top of the water column during their diel migrations (Quetglas et al., 2014; Lebrato et al., 2013; Olivar et al., 2022). As foraging on phytoplankton and small zooplankton is the main reason for the nocturnal migration of intermediate trophic level consumers to the top layers of the water column (Hays, 2003), it is expected that the abundance of their air-breathing predators, such as hard-shelled turtles and striped dolphins, peaks in areas rich in chlorophyll-a or high net primary productivity.

It should be noted, however, that the concentration of chlorophyll-a in the study region peaks in two distinct areas: a high concentration area over the Iberian shelf, due to the fertilization associated with the nutrients transported by freshwater runoff (Bosc et al., 2004), and a secondary peak on the off-shore region of the salinity front that usually lies over the slope of the Iberian Peninsula (Granata et al., 2004; Gordoa et al., 2008; Oguz et al., 2015). The concentration of chlorophyll-a in the latter region is enhanced by upwelling and the meandering of the North Current (Granata et al., 2004), revealed in our analyses by the module of the eastward sea water velocity. In this scenario, the probability that a pixel contains habitat suitable for hard-shelled turtles increased monotonically with the concentration of chlorophyll-a and peaked at negative values of the eastward sea velocity, denoting a strong westward current. Nevertheless, the association between the sightings of striped dolphins and the concentration of chlorophyll-a is non-monotonic, probably because distance to the shore and bathymetry are also very important for this species.

Contrarily, the results presented here do not confirm the concentration of chlorophyll-a as a major environmental driver of the distribution of fin whales at the study area, as reported previously (Panigada et al., 2017a; Tort Castro et al., 2022; Panigada et al., 2024). This is probably because the surveys were conducted during a period of extreme drought (https://www.meteo.cat/climatologia), a scenario in which primary productivity declines and the salinity front approaches the continental shelf due to weak river discharge (Sabates and Olivar, 1996; Gordoa et al., 2008). Krill distribution in the study area is poorly known, but it has been suggested to be associated with the salinity front (Ventero et al., 2019). If so, this may explain why fin whales are usually observed foraging on mid-slope, chlorophyll-rich areas during their northward spring migration across the study area (Panigada et al., 2017a; Tort Castro et al., 2022; Panigada et al., 2024). However, fin whales definitively avoid the continental shelf (Laran and Gannier, 2008; Panigada et al., 2008; Druon et al., 2012) and hence they are unlikely to find suitable foraging grounds rich in krill when the salinity front moves onshore. This hypothesis is supported by the behavior of a few satellite-tracked fin whales in the study area in the spring of 2023, as all of them were in transit across the study area and engaged in area-restricted search only when reaching the Gulf of Lions (Panigada et al., 2024).Accordingly, the concentration of chlorophyll-a is a good predictor of fin whale presence in the study area during the spring only in rainy years, when massive freshwater run-off creates a well-developed, highly productive salinity front over the continental shelf and migrating fins whales engage in foraging. In dry years, fin whales remain associated to the slope of the Iberian Peninsula during the migration, independently from foraging conditions.

The remaining three cetacean species censed in the study area are able to perform much deeper dives than those of sea turtles, striped dolphins and fin whales, reaching at least the uppermost part of the mesopelagic realm in the case of Risso’s dolphins (Arranz et al., 2019; Visser et al., 2021) and the 1000 isobath in the case of Cuvier's beaked whales and sperm whales (Tyack et al., 2006; Watwood et al., 2006). Previous research had already revealed bathymetry as the major determinant of the distribution of these three species, although primary productivity may play also a role in some areas (Gómez de Segura et al., 2006a; Praca et al., 2009; Cañadas et al., 2018; Pirotta et al., 2020; Arcangeli et al., 2023; Chicote et al., 2023). According to the data reported here, the standard deviation of the slope was a major bathymetric feature to determine the distribution of the three species.

The three deep-diving species rely heavily on mesopelagic cephalopods (Rendell and Frantzis, 2016; Santos et al., 2001; Foskolos et al., 2020; Luna et al., 2022), which engage only partially in diel vertical migrations and remain usually below the deep scattering layer and close to the seabed, even at night (Quetglas et al., 2014). In this scenario, only diving regularly over 500 m allows easy access to mesopelagic cephalopods. This lays within the diving capabilities of Cuvier’s beaked whales and sperm whales (Drouot et al., 2004; Watwood et al., 2006; Tyack et al., 2006) and may explain why they lack a clearly defined diel foraging cycle (Giorli et al., 2016). Conversely, the much smaller Risso’s dolphins forage mostly at night, when they perform shallow dives within the epipelagic realm, and it is only in the afternoon when adults engage in deep dives reaching the mesopelagic realm (Giorli et al., 2016; Arranz et al., 2019; Visser et al., 2021). This indicates that Risso’s dolphins are at least partially dependent on the diel vertical migration of their prey. In any case, the results reported here revealed a strong association to areas with a high slope standard deviation (i.e., areas with a heterogeneous slope), although it is unclear why a complex seabed bathymetry may improve access to mesopelagic cephalopods. Previous research indicated their overall scarcity within the Mediterranean Sea and their occurrence only in a few specific sites (Gómez de Segura et al., 2006a; Laran et al., 2017; Chicote et al., 2023; Cañadas et al., 2023). Submarine canyons were identified as critical habitat for the species in the study area in Chicote et al. (2023), although the study suggests a displacement to the more oceanic areas in the 2009–2021 period. Similarly, Azzellino et al. (2016) reported a distributional shift away from the coastal and continental slope habitats and related them to environmental variability, depletion of resources by fisheries and interspecies competition. The habitat suitability models developed here confirmed submarine canyons as highly suitable habitats, but also other regions with complex bathymetry, such as the Spartacus and Brutus seamounts, northwest to Menorca Island.

The capacity of sperm Cuvier's beaked whales to dive at depths greater than 500 meters and up to 1000 meters give them an almost unrestricted access to mesopelagic cephalopods, which may explain why the estimated population biomass of these two species is much larger than that of Risso’s dolphins. Suitable habitat for Cuvier’s beaked whale occurred mostly along the central and northern Iberian slope, whereas the western slope of the Balearic Islands was the area where sperm whales concentrated. Previous studies have suggested that chlorophyll-a, in combination with slope, was one of the major drivers of the distribution of sperm whales in the Mediterranean Sea (Gannier et al., 2002; Praca et al., 2009). However, the highly oligotrophic slope of the Balearic Islands (Bosc et al., 2004) supports one of the largest densities of sperm whales in the Mediterranean Sea (Pirotta et al., 2020; this study), thus suggesting that the concentration of chlorophyll-a is actually non-relevant. This hypothesis is further supported by the negative relationship between habitat suitability and the 3-week-lagged chlorophyll-a concentration, which makes little biological sense. The overall evidence is that chlorophyll-a concentration is not a driver of the distribution of sperm whales within the study area and that the observed correlations between habitat suitability and chlorophyll-a concentration may have arisen from as yet unknown interactions.

Regarding Cuvier’s beaked whales, little is known about their distribution within the Mediterranean Sea, and the available evidence about the suitability of the study area for the species is somewhat contradictory (Cañadas et al., 2018; Arcangeli et al., 2023). Nevertheless, it is worth noting that all the studies conducted to date highlight the slope off the Ebro delta and other parts of the Iberial slope as a highly suitable habitat for the species within the study area (Cañadas et al., 2018; Arcangeli et al., 2023; this study).

In conclusion, the location of suitable habitat patches for epipelagic air-breathing predators (striped dolphins, sea turtles and fin whales) within the Mediterranean Cetacean Migration Corridor is determined largely by the position of the salinity front associated with the Northern Current, whereas the distribution of suitable habitat patches for deep divers (Risso’s dolphins, Cuvier’s beaked whales and sperm whales) is mostly determined by the heterogeneity of the slope. If this hypothesis is correct, dynamic spatial closures might be needed to reduce the potential impact of fishing and maritime traffic on epipelagic predators, whereas static closures might suffice for deep divers. This is probably true also in other geographic areas, but should not be extrapolated without further assessment. On the other hand, RF and GAM yielded similar species distribution models, although sometimes selected different environmental parameters to model de distribution of the same species. This is because they are phenomenological and not mechanistic in nature and underlines the need for combining several modelling approaches to make more robust inferences about the distribution and characteristics of suitable habitats of species.