Areas of spatial overlap between common bottlenose dolphin, recreational boating, and small-scale fishery: management insights from modelling exercises

Study area

The study area (about 400 km²) is in northwest Sardinia, Italy (40.5580 N, 8.3193E, Fig. 1) and includes three coastal protected areas: the Capo Caccia - Isola Piana Marine Protected Area (MPA), the Site of Community Importance (SCI) ‘Capo Caccia e Punta del Giglio’ ITB010042 and the SCI ‘Entroterra e zona costiera tra Bosa, Capo Marargiu e Porto Tangone’ ITB020041. The MPA was established in 2002, mainly for the protection of endemic Mediterranean seagrass (Posidonia oceanica), coralligenous reefs and high priority species, such as the bottlenose dolphin. Human activities (fishing, diving, recreational boating and anchoring) are allowed but regulated in the zones of general protection (Zone B and C), while in Zone A (integral protection) no activity is permitted (with the exception of scientific research authorized by the Management Body). Along the coast the only town is Alghero, whose harbor (the largest on the western coast of the island) hosts 2,200 berths and a fishing fleet mainly composed of small-scale fishing vessels (n = 75; mean length ± SD = 7.18 ± 2.10; “Isola Piana: risultati del monitoraggio 2022 e confronto con le annualitá precedenti”; G La Manna, M Moro Merella, A Ruiu, G Ceccherelli, 2022, unpublished data). The local economy relies on sea-related tourism activities (La Manna et al., 2016); fishing is mainly small scaled and based on traditional practice (“Isola Piana: risultati del monitoraggio 2022 e confronto con le annualitá precedenti”; G La Manna, M Moro Merella, A Ruiu, G Ceccherelli, 2022, unpublished data). Bottlenose dolphin (Tursiops truncatus) is the only species of cetacean regularly present in the area. The local population counts 130 photo-identified individuals, with at least 30% of them showing high site fidelity (La Manna et al., 2023b).

Field data collection

Data on fishing, boating and dolphins were collected from April to September 2022. Systematic standardized surveys were conducted as previously described in La Manna et al. (2020) from two identical 9.7 m motorboats, equipped with 270 hp sterndrive engines. Survey routes followed a haphazard sampling procedure (La Manna, Ronchetti & Sarà, 2016; La Manna et al., 2020) and were planned to homogeneously cover the study area with a generally perpendicular direction with respect to the coast and depth contours. In each survey, boat speed was kept between 6 and 10 kn, to ensure dolphin detection, while two experienced observers scanned the sea surface with both naked eyes and with the help of binoculars. Scans occurred from 9 am to 6 pm, only in days with a visibility of over 3 miles and good sea conditions (sea state 2 Douglas; wind force 2 Beaufort). The position of the research boat was automatically recorded every 30 s using a Garmin GPS chart-plotter. Using binoculars (10 power magnification, 42 mm objective lens), trained observers counted boats, both moored and underway, and fishing buoys encountered within 300 m during the survey, recording their position using GPS. Boats were classified as: recreational inflatable boats, speed boats, cabin cruisers, sail boats, diving boats, fishing boats and tourist ferry boats. Fishing buoys (thus the attached gears) were classified as fishing net or traps, based on their peculiar and distinctive positions (Fig. 2). When the classification was ambiguous the gear was classified as not determined (ND) and excluded from the analysis. In case of sighting, dolphins were approached cautiously (to avoid disturbance) within 10–50 m and dolphin group size was estimated by two independent observers. The geographic position of the group (defined as all individuals within visual range that were in apparent association, engaged in the same activity or moving in the same direction; Shane, 1990) was collected using a Garmin GPS chart-plotter at the beginning and at the end of the sighting. Only the beginning sighting points were considered as dolphin occurrence records in the analysis. During each dolphin sighting, a photo-identification data collection was also performed to allow the identification of each dolphin based on the presence of natural marks, (nicks, scars, or skin pigmentations on the dorsal fin;Würsig & Jefferson, 1990). Based on photo-identification analysis, when the same group of individuals was observed several times in the same day, only the first sighting was used in the modelling exercise to avoid the non-independence and temporal autocorrelation of sightings.

Environmental predictors

We incorporated data on environmental variables, dolphin, fishing gear, and boat occurrence records into a Geographic Information System (GIS; software ArcMap 10.8) using the World Geodetic System 1984 (WGS84) and the Universal Transverse Mercator (UTM) 32 N projection.

The following environmental variables were selected to model dolphin habitat suitability, based on previous studies on cetacean (Azzellino, Panigada & Lanfredi, 2012; Marini, Fossa & Paoli, 2014; Bonizzoni et al., 2019; La Manna, Ronchetti & Sarà, 2016; La Manna et al., 2020): water depth (m), mean sea surface temperature (SSTm), SST range (SSTr, as difference between the maximum and the minimum sea surface temperature), mean chlorophyll concentration (Chl-a - mg/m³), seafloor slope (degree), distance to the coast (m), aspect (downslope direction) and seabed habitat type (Table 1, Suppl. Mat. 2). All data were pooled into grids with a resolution of 250 m and averaged over two periods (spring: from April to June; summer: from July to September) and the whole period (spr-sum). We used ‘Raster Correlations and Summary Statistics’ tools of the plug-in SDMtoolbox 2.5 for ArcGis to create a correlation matrix of the Pearson correlation coefficients (r) between the environmental predictors (Brown, Bennett & French, 2017). Since correlation (r > 0.8) was found between distance to the coast, chlorophyll concentration and depth (Suppl. Mat. 1), only depth was used in the modelling exercises (Kramer-Schadt, Niedballa & Pilgrim, 2013).

Model generation and validation

We modeled boat, fishing gear, and dolphin occurrence records at different temporal scales by Maximum Entropy (MaxEnt version 3.4.1, Phillips, Dudík & Schapire, 2020), one of the most reliable tools (Elith, Graham & Anderson, 2006; Elith, Phillips & Hastie, 2011) when presence-only data (Guisan & Thuiller, 2005) are available and the number of sightings is low (Elith, Graham & Anderson, 2006; Valavi et al., 2022). MaxEnt estimates a probability distribution (i.e., relative likelihood of presence; McClellan et al., 2014) for a species by contrasting occurrence data with background data, rather than true absence data (Thorne et al., 2012). Thus, MaxEnt does not assume that the absence precludes the likelihood of occurrence, and this aspect is particularly relevant in case of highly elusive species, such as cetaceans (La Manna, Ronchetti & Sarà, 2016; La Manna et al., 2016; La Manna et al., 2020). MaxEnt was preferred to other Species Distribution Model algorithms due to its general higher performance with presence-only data, in comparison with other models (Valavi et al., 2022), and because it allowed to compare the dolphin habitat suitability predicted in 2022 with that relative to the period 2015–2018 obtained with the same modeling approach (La Manna et al., 2020).

Boating and fishing activities were modelled by MaxEnt since they vary temporally and spatially as a function of human behavior (Cummins, Gault & O’Mahony, 2008; La Manna et al., 2020). Thus, we modelled the likelihood of fishing gear and boat presence using the occurrence records of all boats and fishing buoys (distinguished in traps and fishing nets) recorded during the surveys as presence data and the raster surface of depth, slope, aspect, and seabed habitat types as explanatory variables (Table 2; Suppl. Mat. 2). These environmental predictors were chosen since seabed morphology and habitats may affect both fisher and boater behavior. For modeling the likelihood of fishing gear presence, we also considered SSTm and SSTr, since temperature is among the main driver of fish distribution (Azzurro et al., 2019), and distance to harbor (as categorical variable), since the distance between the harbor and the fishing site could influence fishers’ choice (Table 2). Other predictors, such as tidal currents and wave height, that have been used in other fishing modelling exercises (van der Reijden et al., 2018), were not considered here because of their irrelevance in the local (small scaled) context.

For modeling the likelihood of dolphin presence, we considered four continuous environmental variables (SSTm, SSTr, depth and slope), two categorical environmental variables (aspect and seabed habitat type) and two anthropogenic predictors (the likelihood of boat and fishing net presence modeled by MaxEnt). Since correlation was found between the likelihood of fishing net and trap presence, only the former was used in modelling dolphin presence (Suppl. Mat. 1).

Following La Manna et al. (2020), for all models, MaxEnt settings were chosen in relation to the specific questions of the study and data limitations (Merow, Smith & Silander, 2013): (i) the cloglog output since it is most appropriate for estimating likelihood of presence (Phillips, Dudík & Schapire, 2020), (ii) hinge features (Phillips & Dudík, 2008), (iii) default regularization parameters (0.5), and (iv) 10-fold cross validation to assess the average behavior of the algorithms randomly partitioning the occurrence data in ten random subsets (Phillips, Anderson & Schapire, 2006). The geographical extent of the models coincides with the monitored area that includes the typical and known habitat of bottlenose dolphins (Bearzi, Fortuna & Reeves, 2012). Thus, we used all the background sites available (5,358) to increase the predictive performance (Phillips & Dudík, 2008). Furthermore, since background points were generated from the same environmental space as the presence locations, bias files, useful to fine-tune the selection of background points in MaxEnt and to account for sampling bias (Brown, 2014), were not used.

We ran a jackknife analysis to estimate the contribution of each variable to the MaxEnt run, obtaining alternative estimates of variable importance for our models. The performance of each MaxEnt model was assessed using the AUC (area under the receiver operating characteristic curve; Phillips, Anderson & Schapire, 2006), a threshold-independent metric of overall accuracy. The AUC provides an adequate evaluation of model performance (Phillips, Anderson & Schapire, 2006), even if, in the case of presence-only distribution models, AUC will compare presences with background points and cannot be considered a perfect measure of model accuracy (Lobo, Jiménez-Valverde & Hortal, 2010; Merow, Smith & Silander, 2013). AUC values range between 0 and 1: models with moderate to-good discrimination ability correspond to AUC higher than 0.7 (Swets, 1988). Following a procedure already applied in La Manna et al. (2020), the model robustness was also evaluated calculating (i) the test-AUC standard deviation (SD) and the difference between the train-AUC values (using all presences) and the mean test-AUC values, since lower are the test-AUC SD and the difference between the train-AUC and mean test-AUC values higher is the model robustness (Herkt et al., 2016).

Using the cloglog output of MaxEnt (Phillips, Dudík & Schapire, 2020), we produced maps for each season (spring and summer) and the whole period (spr-sum) relative to the likelihood of boat, fishing net and dolphin presence.

To visualize the overlapping area between dolphins and boats and between dolphins and fishing nets, the intersections of their representative areas were calculated, for both seasons and the whole period, using as representative areas those with likelihood of presence >0.6. This value corresponds to the threshold which balances sensitivity (the change of correctly identifying suitable areas) and specificity (the change of correctly identifying not suitable areas) (Liu, White & Newell, 2013). Finally, two percent area overlap, one between dolphins and boats (PAO_d,b) and the other between dolphins and fishing nets (PAO_d,f), were calculated following Atwood & Weeks (2003): ${\displaystyle PA{O}_{\mathrm{d,b}}=\left(A_{\mathrm{d,b}}/A_{\mathrm{d}}\right)\times {\left(A_{\mathrm{d,b}}/A_{\mathrm{f}}\right)}^{\mathrm{0,5}}}$ ${\displaystyle PA{O}_{\mathrm{d,f}}=\left(A_{\mathrm{d,f}}/A_{\mathrm{d}}\right)\times {\left(A_{\mathrm{d,f}}/A_{\mathrm{f}}\right)}^{\mathrm{0,5}}}$ where A_d,b isthe overlap area between dolphins and boats, A_d,f isthe overlap area between dolphins and fishing nets, A_d is the representative area for dolphins, A_b is the representative area for boats, and A_f is the representative areas for fishing nets.

Boat and fishing net presence

Between April and September, a total of 2,973 boat occurrence records were counted, 657 in spring and 2,316 in summer. The accuracy of MaxEnt in predicting the likelihood of boat presence was higher than 0.75 in any investigated period (Table 2). The most important contributor in predicting the likelihood of boat presence was depth, followed by aspect (Table 2). The probability of finding boats was higher between -60 and -15 m of depth and similar for all aspect classes, except for the classes 2 (NW) and 7 (W), the orientation most exposed to the dominant local wind (Suppl. Mat. 2). Overall, the highest likelihood of boat presence was along the north-western part of the Alghero coast and inside the boundaries of the MPA/SCI (Fig. 3), especially in summer.

A total of 409 fishing net occurrence records were counted, 175 in spring and 234 in summer. The accuracy of MaxEnt in predicting the likelihood of fishing net presence was higher than 0.8 in any investigated period (Table 2). The most important contributors in predicting the likelihood of fishing net presence were depth (in summer), SSTm, distance to harbor and seabed habitat type (in spring) (Table 2). In spring, the probability of finding fishing nets was higher in the habitat “Mediterranean communities of muddy detritic bottoms” and “Posidonia bed”, and within 4 km to the harbor (Suppl. Mat. 3), while in summer the probability of finding fishing nets was higher between -60 and -20 m of depth and between 8 and 12 km from the harbor. In any period, the increase of SSTm decreased the probability of finding fishing nets. Overall, the likelihood of fishing net presence was highest along the coast closest to Alghero, especially in spring, but also in the areas inside and surrounding the MPA/SCI, especially in summer (Fig. 3).

Dolphin presence

Across all periods, 110 dolphin groups were sighted in about 5,650 km of navigation and 163 surveys. All MaxEnt models obtained AUC higher than 0.75, indicating high accuracy in predicting dolphin presence (Table 2). Further, the overall model robustness can be inferred by the small difference between train-AUC and mean test-AUC values (Table 2).

The most suitable areas for dolphins are inside the Alghero Bay, both along its western and southern coasts, and only partially overlap with the boundaries of the MPA/SCI (Fig. 3). The most important contributing variables to the dolphin habitat suitability models were the likelihood of fishing net presence, seabed habitat type, aspect, SSTm and the likelihood of boat presence (Table 2). Particularly, the probability of finding dolphins: (i) increased as the likelihood of fishing net presence increased, in all seasons; (ii) changed with the seabed habitat types and aspect classes, depending on the season; (iii) decreased when the SSTm increased in summer; iv) decreased when the likelihood of boat presence was higher than 0.5 in spring (Fig. 4).

The area with the likelihood of dolphin presence higher than 0.6 extended for about 58 km², ranging from 49 km² in spring to 61 km² in summer (Table 3). This area largely overlapped with both the areas with the highest likelihood of fishing net and boat presence, mainly in summer (Table 3; Fig. 5).

Study limitations

Some ecologically relevant results were obtained with this study, although not all the potential factors influencing dolphin habitat suitability were included in the models. Thus, the limited number of predictors constrains the interpretation of the results within the range of the investigated environmental conditions (La Manna, Ronchetti & Sarà, 2016; La Manna et al., 2016; La Manna et al., 2020). For example, although fishing and SST can be considered good proxies of prey availability, the model predictability would be certainly improved including data on fish abundance and distribution (MacLeod et al., 2014). Further, other aspects remain to be clarified. For example, the distribution of dolphins during the winter months, when sea-based tourist activities cease and fishing reduce due to adverse weather conditions, as well as the influence of different behavioral states (feeding, socializing, resting) on the selection of habitats (Giannoulaki et al., 2017) should be investigated. Finally, considering that marine mammals respond to climate change (Silber Gregory et al., 2017), and early indications on the effect of temperature on the studied population have already been found (La Manna et al., 2023a; La Manna et al., 2023b), further effort should be done to anticipate changes in occurrence, distribution, and relative abundance of this population under climate change scenarios.