Harnessing citizen science for marine conservation in Malta: a comparative analysis of GAM and MaxEnt models in bottlenose dolphin habitat mapping

Introduction

Understanding species distribution and habitat use is fundamental for developing effective conservation strategies (Rodríguez et al., 2007; Guisan et al., 2013). However, assessing these aspects can be particularly challenging for highly mobile species like cetaceans (Fernandez, Sillero & Yesson, 2022). In this context, species distribution models (SDMs) offer a solid foundation for gaining valuable ecological insights, particularly in data-deficient regions (Redfern et al., 2006; Fiedler et al., 2018; Fernandez et al., 2021). By evaluating the relationships between cetacean populations and the environment, as well as anthropogenic factors, SDMs contribute to a more comprehensive understanding of species ecology and potential conservation priorities (Rodríguez et al., 2007; Guisan et al., 2013; Marshall, Glegg & Howell, 2014; Pace, Tizzi & Mussi, 2015; Giralt Paradell, Díaz López & Methion, 2019).

The ability of SDMs to assess species distributional ranges via predictive modeling (Anderson, Lew & Peterson, 2003; Pitchford et al., 2016) or to identify the environmental drivers of distribution through explanation modeling (Azzellino et al., 2008; La Manna et al., 2023b) has led to an increase in their use in recent years for cetacean species (Pasanisi et al., 2024). When used together, predictive and explanation modeling provides information on the distribution of habitats that are suitable for species survival (Hirzel & Le Lay, 2008). Within this framework, habitat suitability models serve as practical tools to estimate the ecological conditions and areas within which a species is most likely to occur (Guisan & Thuiller, 2005).

Among the most widely adopted modeling techniques are machine learning algorithms such as maximum entropy (MaxEnt) models (Melo-Merino, Reyes-Bonilla & Lira-Noriega, 2020), which estimate the habitat suitability from presence-only data using the principle of maximum entropy (Phillips, Anderson & Schapire, 2006), and statistical approaches such as generalized additive models (GAMs), which model non-linear relationships between species occurrence (e.g., presence/absence) and environmental factors (Hastie & Tibshirani, 1986). Both methods enable researchers to identify critical conservation areas and mitigate potential threats by analyzing species’ habitat preferences (Rodríguez et al., 2007; Guisan et al., 2013). Furthermore, they help predict how marine species respond to environmental changes, contributing to the development of targeted conservation strategies (Giralt Paradell, Díaz López & Methion, 2019; Díaz López & Methion, 2024).

The performance and reliability of these models are inherently dependent on several factors, including model parametrization (Elith et al., 2006), the ecological relevance and statistical independence of environmental predictors (Guisan & Zimmermann, 2000; Dormann et al., 2013), the spatial and temporal scale at which models are implemented (Elith & Leathwick, 2009) and the overall quality, completeness, and bias of species occurrence records (Hernandez et al., 2006; Araújo & Guisan, 2006; Merow, Smith & Silander, 2013).

Dedicated surveys, although considered robust, are often resource-intensive and limited by logistical and financial constraints, resulting in incomplete spatial coverage and potential underrepresentation of species distribution (Evans & Hammond, 2004; Meyer et al., 2015). Recent studies have demonstrated that non-traditional data sources, such as opportunistic and citizen science data (Giovos et al., 2016; Pace et al., 2019; Robbins, Babey & Embling, 2020), offer a cost-effective approach to addressing specific research challenges, facilitating the generation of estimates of cetacean distribution (Fernandez et al., 2021). In the study of species inhabiting dynamic environments, non-traditional data sources offer several advantages over conventional survey methodologies. For instance, they can enhance spatial and temporal coverage and reduce logistical constraints, as they often do not require specialized equipment or dedicated field effort (Robbins, Babey & Embling, 2020; Corr et al., 2024). Nevertheless, the inherent biases and variability associated with opportunistic data necessitate careful quality assessment and data integration techniques to ensure their effective use in large-scale modeling efforts (Isaac et al., 2020; Martino et al., 2021).

Given these challenges, the integration of SDMs with non-traditional data sources presents a promising approach for studying cetaceans in regions where systematic monitoring is limited. In Maltese waters, the common bottlenose dolphin (Tursiops truncatus) represents one of the most frequently observed cetacean species (Notarbartolo Di Sciara, 2002) and plays an important ecological role as a top predator within the marine ecosystem. Bottlenose dolphins are protected under a framework of national, regional and international legal frameworks. At the national level, all cetacean species are protected through the Flora, Fauna and Natural Habitats Protection Regulations (S.L. 549.44), which transpose the European Union (EU) Habitats Directive 92/43/EEC into Maltese law. In particular, the bottlenose dolphin is listed in Annex II of the Directive, which requires the designation of Special Areas of Conservation (SACs) forming part of the Natura 2000 network, and in Annex IV as a species of community interest in need of strict protection.

Despite its protected status, the limited availability of systematic data on bottlenose dolphins in Maltese waters presents challenges for fully supporting conservation planning. Previous studies in the region have suggested a coastal distribution and spatial association between dolphin presence and anthropogenic features such as aquaculture sites, which may influence foraging behavior and local prey availability (Laspina, Terribile & Said, 2022; Soster et al., 2025). Recent modeling efforts have highlighted potential gaps in the spatial alignment between currently designated Natura 2000 sites and areas predicted to be of high habitat suitability for bottlenose dolphins around the Maltese archipelago (Soster et al., 2025). Specifically, although SACs have been designated for bottlenose dolphin conservation in Malta, these mainly encompass offshore regions, while key coastal regions remain outside protected boundaries. This highlights the need for further in-depth investigation of bottlenose dolphin spatial ecology and habitat use in these areas, particularly through temporally and spatially specific studies that can build upon existing baseline knowledge. Moreover, dolphins in the region are exposed to a range of human pressures, including maritime traffic, aquaculture, and unregulated recreational activities (Said et al., 2017; Filletti et al., 2023; Mizzi et al., 2024; Soster et al., 2025). They are additionally subject to environmental pressures, such as marine heatwaves (Garrabou et al., 2022), which require further investigation to be adequately addressed under existing management frameworks. These conservation gaps highlight the urgent need for spatially explicit data on dolphin distribution and habitat use to inform evidence-based conservation strategies in Malta’s dynamic and heavily used coastal waters.

In light of these gaps, this case study aims to improve understanding of bottlenose dolphin habitat suitability using presence-only data from opportunistic research platforms and citizen science. We apply GAM and MaxEnt approaches within a heavily impacted coastal SAC that, despite its designation, does not explicitly include bottlenose dolphins as a target species. While this study is conservation-focused and not a methodological comparison of modeling techniques, the inclusion of both models allows for a more nuanced interpretation of species–habitat relationships and enhances stakeholder confidence in the results. By analyzing and comparing the outputs, we identify summer habitat suitability patterns and assess the influence of environmental and anthropogenic drivers. The ultimate long-term goal of this work is to support evidence-based spatial planning and inform the potential inclusion of bottlenose dolphins within the site’s conservation framework.

Materials & Methods

Study area

The study area covers 159 km² and is situated on the northeastern side of the Maltese archipelago, extending from central Malta to the northern coast of Gozo, with a maximum depth of 100 m (Fig. 1). The area includes the Natura 2000 site Żona fil-Baħar bejn Il-Ponta ta ‘San Dimitri (Għawdex)u Il-Qaliet, designated as a SAC for the protection of the Maltese top-shell (Steromphala nivosa) and the loggerhead turtle (Caretta caretta). Although bottlenose dolphins are known to occur in the area, they are not currently listed among the protected species for this site (EUNIS, 2025). This coastal SAC hosts a wide range of important habitats, such as Posidonia oceanica meadows, maërl beds, and Cymodocea nodosa meadows. However, this site is affected by numerous human activities, including four main finfish aquaculture sites and one bunkering area for large vessels (Fig. 1). Furthermore, pleasure boating is especially prevalent, driven by the high concentration of marinas and the region’s appeal as a tourist destination.

Results

Habitat suitability analysis with GAM

The GAM showed strong predictive performance, with a mean AUC of 0.86 (SE ± 0.02) under 10-fold cross-validation, indicating a high level of discriminatory power (Swets, 1988). All four explanatory variables included in the model were found to have statistically significant relationships with bottlenose dolphin occurrence. In terms of variable importance, the proportion of deviance explained was highest for chlorophyll-a (0.79) and distance to aquaculture (0.65), suggesting these were the strongest drivers of dolphin habitat suitability (Table 2). Although slope and SST temperature anomaly were statistically significant, they contributed less to the overall deviance explained (0.20 and 0.16, respectively), indicating a weaker influence on habitat suitability patterns.

The model predicted an increased probability of occurrence with increasing values of slope and chlorophyll-a and a decreased probability of occurrence with increasing values of SST anomaly and distance to aquaculture facilities (Fig. 3).

The mean prediction map generated across 1,000 iterations showed that the highest habitat suitability for bottlenose dolphins was concentrated around Mellieha Bay and St Paul’s Bay. Additional areas of moderate suitability were identified in the Comino Channel, in proximity to the bunkering area and along the north-eastern coast of Gozo (Fig. 4). The uncertainty associated with the spatial predictions, represented by the coefficient of variation across, was generally low within the borders of the SAC but higher at the outer edges of the survey region.

Table 2:

Relative importance of explanatory variables in the generalised additive model (GAM).

The sum of the proportion of the total deviance explained exceeds 1 because the method penalises more complex models.

Variable	Deviance explained	Proportion of total deviance explained*
Chlorophyll-a	31.4%	0.79
Distance to aquaculture	26%	0.65
Slope	7.89%	0.20
SST anomaly	6.36%	0.16

DOI: 10.7717/peerj.19804/table-2

Notes:

*The sum of the proportion of the total deviance explained exceeds 1 because the method penalises more complex models.

Habitat suitability analysis with MaxEnt

The MaxEnt model demonstrated strong predictive performance (Swets, 1988), with a mean AUC of 0.89 (SE ± 0.0001) under 10-fold cross-validation. The regularized training gain was 1.23, compared to an unregularized gain of 1.18, indicating that the application of regularization successfully reduced overfitting to a small extent.

Chlorophyll-a was the most influential predictor in the model, contributing 61.2% of the total model gain and 25.8% of the permutation importance (Table 3). Distance to aquaculture was the second most influential variable, contributing 20.1% to model gain and 53.3% to permutation importance.

The model showed a marked decrease in habitat suitability with increasing distance from aquaculture facilities, particularly within the first 5,000 m. Habitat suitability increased exponentially with higher chlorophyll-a concentrations (Fig. 5). Slope and SST anomaly contributed less than 15% cumulatively to the model’s predictive power. However, slope showed a higher permutation importance (20.4%), suggesting a moderate influence on model sensitivity. The response curve indicated that suitability peaked at slope values between 2.5° and 3°, declining on either side. Sea surface temperature anomaly exhibited a nearly linear negative relationship with suitability, but its overall contribution to the model was minimal (Fig. 5).

As with the GAM, the most suitable areas for bottlenose dolphins were the waters surrounding Mellieha Bay and Saint Paul’s Bay. The MaxEnt model predicted high suitability only along the Maltese coast of the Comino Channel and a moderate suitability in the shallow bunkering area (Fig. 6). The coefficient of variation was highest near the edges of the survey area, and higher uncertainty was also observed within the SAC, especially along the central coastal waters of Gozo.

Table 3:

Percent contribution and permutation importance for the Maximum Entropy Model (MaxEnt).

Predictor	Percent contribution of predictors	Permutation importance
Chlorophyll-a	61.2	25.8
Distance to aquaculture	20.1	53.3
Slope	10.4	20.4
SST anomaly	2.3	0.5

DOI: 10.7717/peerj.19804/table-3

Discussion

Effective conservation planning relies on identifying areas that are important for species survival and on understanding how environmental factors influence their distribution. SDMs are widely used tools for obtaining this information, particularly in data-limited contexts (Elith & Leathwick, 2009; Guisan et al., 2013). This case study identifies core suitable habitats for a top marine predator by linking occurrence records, obtained with the support of citizen science, to key environmental variables. Through the application of a GAM and a MaxEnt model, summer habitat suitability maps were generated for the bottlenose dolphin in a relatively understudied area of Malta, offering a practical tool to support decision-makers in the formulation of conservation strategies.

This study examined the bottlenose dolphin summer habitat suitability, revealing that the northeastern coastal waters of the archipelago represent an important summer habitat for the species. In addition to being in line with previous broader research in the region (Soster et al., 2025), this study provides a finer-scale perspective by adding spatial detail and seasonal context to earlier habitat suitability assessments. Both the GAM and the MaxEnt models used in this research identified the area between Mellieha Bay and St Paul’s Bay as a region of high suitability, suggesting that this portion of the study area may represent a key habitat for the species in summer. This area coincides with shallow, productive waters and proximity to inshore aquaculture sites, factors previously associated with higher dolphin presence in other Mediterranean regions, where similar environmental and anthropogenic features have been linked to increased dolphin presence (Díaz López, 2012; Pace et al., 2019; Gnone et al., 2022; La Manna et al., 2023a; Bellingeri et al., 2025). Moderate suitability was also found across the broader study area, with particular interest in the Comino Channel, but with some degree of spatial variability between the two models.

To better understand the environmental drivers behind these patterns, the relative importance of predictor variables for both models was examined. While GAM and MaxEnt differ in their analytical frameworks, both models’ outputs revealed a comparable pattern in the importance of environmental variables shaping bottlenose dolphin habitat preference. Particularly, both models identified chlorophyll-a and distance to aquaculture as the most important predictors of habitat suitability, while slope and SST anomaly contributed less significantly. However, the relative contribution of these variables differed between the two approaches. In the GAM, chlorophyll-a explained the highest proportion of deviance, followed closely by distance to aquaculture. In contrast, MaxEnt revealed distance to aquaculture as the most influential predictor based on its permutation importance, with chlorophyll-a as the second most important variable. These differences reflect how each model handles variable interactions (Guisan & Zimmermann, 2000; Elith et al., 2011), but the consistency in identifying chlorophyll-a and distance to aquaculture as the top predictors highlights their ecological relevance to this population.

To further clarify these ecological patterns, the response curves offer information on how particular environmental factors affect habitat suitability. Higher chlorophyll-a concentrations showed a positive relationship with habitat suitability in both models, likely reflecting higher primary productivity and, by extension, greater prey availability (Methion et al., 2023). Chlorophyll-a has frequently been used as a proxy for such favorable foraging conditions, given its correlation with the abundance of zooplankton and small pelagic fish (Torres, Read & Halpin, 2008). Indeed, its importance as a predictor of dolphin occurrence and habitat use has been demonstrated in other regions of the Mediterranean (La Manna, Ronchetti & Sarà, 2016; Giannoulaki et al., 2017; Karamitros et al., 2020). In this study, chlorophyll-a levels were likely influenced by both natural productivity and anthropogenic input from aquaculture (Díaz López & Methion, 2017), as chlorophyll-a concentrations were consistently higher in areas surrounding aquaculture sites during the summer months (Fig. S1), contributing to their predictive strength in both models. Additionally, bottlenose dolphins are highly opportunistic predators with a well-documented behavioral plasticity (Reynolds, Wells & Eide, 2013), often concentrating their foraging activity in areas where prey is predictably abundant. Suitability also declined with increasing distance from aquaculture sites, suggesting that these areas may provide foraging opportunities by promoting local prey aggregations (Díaz López, 2012; Díaz López, 2017; Díaz López & Methion, 2017). However, this does not imply that the area would be unsuitable in the absence of aquaculture; rather, the presence of these sites appears to increase habitat use within an ecologically favorable region. Finally, while the observed negative correlation between bottlenose dolphin presence and distance to aquaculture facilities improves our ability to predict their habitat preferences, this aggregation near fish farms may pose conservation concerns, as previous studies have associated such clustering with increased risks of vessel collisions, entanglement, habitat degradation, and exposure to elevated noise and light pollution (Guisan & Zimmermann, 2000; Martino et al., 2021).

While chlorophyll-a and distance to aquaculture were the strongest predictors in both models, slope and SST anomaly also contributed to explaining dolphin habitat suitability, albeit to a lesser extent. In the GAM, slope showed a significant but weaker relationship, with habitat suitability peaking at intermediate values. Similarly, the MaxEnt response curve indicated that suitability decreased on either side of an optimal slope range, suggesting a preference for specific seabed structures. Both models indicated a preference for areas characterized by gentle slopes (∼2.5–3°), indicating that bottlenose dolphins in the study area are associated with mildly inclined rather than very flat or steep seabed. This is consistent with previous findings suggesting that moderate bottom inclinations may support preferred foraging or transit zones, particularly in coastal environments where bathymetric features influence prey availability (Cañadas, Sagarminaga & Garcı’a-Tiscar, 2002; Gnone et al., 2022).

Although SST anomaly was the weakest predictor in both models, the response curves indicated a consistent negative relationship between increasing temperature anomalies and habitat suitability. This suggests that bottlenose dolphins may tend to avoid areas experiencing greater thermal deviation, potentially due to indirect effects on prey availability or ecosystem stability (Wernberg et al., 2013). In the Mediterranean, SST anomalies have been associated with shifts in the abundance and distribution of demersal and pelagic fish species (Sabatés et al., 2006; Hidalgo et al., 2011), which could in turn reduce the attractiveness of the areas affected to predators. Given that the study only covered summer months, the role of thermal variability might be underestimated, and further investigation across seasons would be needed to better understand its influence. This is of particular importance, in light of the potential consequences of extreme weather events such as marine heatwaves that are becoming more frequent in the Mediterranean (Darmaraki et al., 2019; Wild et al., 2019).

One important finding of this study is the spatial agreement shown by the two models in predicting areas of high habitat suitability, particularly in the waters surrounding Mellieha Bay and St Paul’s Bay. However, a certain difference between the two modeling approaches was observed. Particularly, the GAM predicted a broader extent of high suitability, including the Comino Channel, as well as moderate suitability across most of the study area. In contrast, the MaxEnt model identified a more spatially restricted area of high suitability in the channel, primarily concentrated in Marfa Bay, with moderate suitability extending across the bunkering area. Additionally, while the coefficient of variation was higher around the SAC boundaries for both models, the MaxEnt predictions exhibited greater uncertainty within the SAC itself, particularly around the northern coastal waters of Gozo relative to the overall coefficient of variation. These differences likely reflect the methodological dissimilarities between the models and their respective sensitivities to the spatial distribution and structure of the input data. GAM’s use of pseudo-absence data and smooth functions may have allowed it to detect localized patterns of suitability, particularly if species-environment relationships in this area are non-linear or influenced by smaller gradients (Guisan, Edwards & Hastie, 2002; Wood, 2017). In contrast, MaxEnt’s reliance on background sampling and global feature fitting may have reduced sensitivity to weak or spatially clustered signals in regions with lower observation density (Elith et al., 2011; Merow, Smith & Silander, 2013). Interestingly, while MaxEnt’s outputs have been found to be only similar to those of GAMs when background points are drawn from observed absences (Fiedler et al., 2018), these results show spatial agreement despite the use of pseudo-absences and background data, suggesting a robust ecological indicator captured by both modeling approaches. These model-specific differences highlight the value of employing complementary modeling techniques, as they can reveal different aspects of species-habitat relationships and strengthen the ecological interpretation of suitability maps (Guisan & Zimmermann, 2000; Elith & Graham, 2009).

Given its significance in conservation applications of SDMs (Guisan et al., 2013), it is important to highlight the low degree of uncertainty found in the central core area identified between Mellieha Bay and St. Paul’s Bay. While some uncertainty is inherently irreducible (Regan et al., 2005), the low uncertainty and strong spatial agreement of highly suitable areas, together with the high predictive performance of both models (AUC > 0.85), reinforce confidence in the identified ecological patterns. Making the uncertainty explicit also helps stakeholders understand the level of confidence they can place in the prediction (Regan et al., 2005). The highly suitable area found by both models should therefore be treated as a conservation priority, while the broader GAM-predicted moderately suitable areas may inform future monitoring efforts. In data-limited contexts, this complementarity can provide a more holistic and robust foundation for decision-making.

Incorporating citizen science data was a key aspect of this study, enabling a nearly threefold increase in presence records as well as expanding the number of sightings in areas with less survey coverage. In contrast, models based solely on research data predicted low habitat suitability in the Comino Channel, despite frequent dolphin sightings reported by citizen scientists (Fig. 2). The reduced coefficient of variation in both the GAM and MaxEnt models using the full dataset further underscored the value of including opportunistic observations, especially when systematic survey effort is limited. These findings align with previous research showing that appropriately incorporating citizen science and opportunistic data can significantly enhance species distribution models (Tiago, Pereira & Capinha, 2017; Coxen et al., 2017; Matutini et al., 2021). Our results contribute to the growing recognition of non-traditional data sources as valuable tools for contributing to cost-effective long-term monitoring and the generation of reliable habitat suitability predictions. These predictions help govern conservation, particularly in marine environments where systematic data collection is often constrained (Giovos et al., 2016; Pace et al., 2019).

Protecting bottlenose dolphins in Maltese waters effectively calls for the inclusion of species–habitat knowledge into site-based and pressure-based management approaches (Reisinger, Johnson & Friedlaender, 2022). This study adopts a conservation-oriented perspective and identifies key summer habitats which, although located within an existing SAC, are not currently designated for the protection of bottlenose dolphins. The strong agreement between habitat suitability maps highlights consistent ecological patterns and supports spatial planning, local mitigation, and the potential inclusion of bottlenose dolphins in the site’s conservation objectives. Furthermore, the integration of citizen science data demonstrates how participatory monitoring expands spatial coverage and enhances model reliability in under-surveyed areas. Employing complementary modeling approaches, such as MaxEnt and GAMs, further strengthens the ecological interpretation of habitat use and provides evidence for adaptive management in data-limited marine environments.

Conclusions

This study applied a comparative modeling approach using generalized additive models and MaxEnt to assess summer habitat suitability for bottlenose dolphins in a coastal Special Area of Conservation (SAC) in Malta. The models integrated citizen science and opportunistic data with environmental and anthropogenic variables to identify key drivers of dolphin habitat suitability and highlight important habitats within a data-limited context.

Both modeling approaches consistently identified shallow, nearshore regions, particularly around Mellieha Bay and St Paul’s Bay, as key summer habitats for bottlenose dolphins. Chlorophyll-a concentration and proximity to aquaculture sites emerged as the strongest predictors, while slope and sea surface temperature anomaly played a secondary role. Despite some methodological differences between the models, the observed spatial agreement strengthens confidence in the ecological relevance of these findings.

By combining traditional ecological knowledge with citizen science contributions, this work demonstrates a scalable and cost-effective approach for habitat assessment, especially valuable in regions where systematic monitoring is limited. The results contribute to filling critical knowledge gaps on dolphin spatial ecology in Malta’s coastal waters and offer science-based support for potential updates to conservation planning frameworks, including the consideration of bottlenose dolphins as a qualifying species within the SAC.

However, this study is not without limitations. The research focused on a single summer season and was based on presence-only data, limiting its ability to capture interannual variability and broader seasonal dynamics. Additionally, while sampling bias was addressed, the use of opportunistic data remains susceptible to uneven spatial effort and reporting.

Future research should prioritize increasing the temporal resolution of data collection across multiple seasons and years, improving systematic survey coverage, and further investigating the ecological linkages between aquaculture, productivity hotspots, and dolphin habitat use. Exploring these dynamics under different climatic scenarios, including marine heatwaves, may also provide critical insights for adaptive conservation management.

In summary, this case study contributes to the growing body of evidence supporting the integration of SDMs and participatory data collection into marine conservation planning. The approach presented here offers a valuable tool for informing spatial management decisions aimed at balancing species protection with human activities.

Supplemental Information