An interview-based approach to assess sea turtle bycatch in Italian waters

Material and Methods

Study area

The study was carried out in Italy (central Mediterranean Sea, Fig. 1), which encompasses loggerhead migration routes (Sicily channel and Ionian Sea; Bentivegna, 2002), foraging areas (Adriatic Sea; Casale et al., 2012) and stable nesting rookery in the Ionian Sea (Mingozzi et al., 2007; Garofalo et al., 2009). Data are presented by Geographical Sub-Areas (GSAs) of the General Fisheries Commission for the Mediterranean comprising the Italian coastline, assuming that the bycatch relating to each fishing method and environmental condition in each GSA are sufficiently similar to provide a homogenous bycatch amount. The data from each GSA were then divided by fishing method and season.

Interview survey design

More than 30 interviewers from 9 institutions from different parts of Italy, including a research body, a regional authority, 2 non-profit organizations, 2 private organizations, 2 marine protected areas and 1 national park, participated in the survey. In total 453 interviews were conducted in 105 Italian fishing harbours, covering all Italian GSAs and about 98% of the Italian coast length (total 7,458 km). About 6% of the entire Italian fleet was interviewed. Sampling distribution across geographical areas was generally in line with the distribution of the fishing fleet, even though the consistency of data collection and reporting varied across GSAs and fishing methods.

Face to face interviews were organized directly in the harbour, on board fishing vessels and during fishermen’s associations meetings. A single questionnaire was used for all fisheries and fishing methods.

The interview consisted of 5 sections: (1) ‘background information’ involved questions related to the fisherman’s experience, fishing gear used and fishing grounds; (2) ‘frequency of turtle encounters and fishermen’s behaviour’ involved questions on the number of sea turtles caught per season in 2014 and on their management upon capture; (3) ‘suggestions to reduce turtle bycatch and knowledge of bycatch reducer devices’ asked how the interviewee thought that turtle bycatch could be reduced and for his opinion on the adoption of Bycatch Reducer Devices (BRDs) and gear modifications (such as circle hooks for longlines, Turtle Excluder Devices for trawl nets or light deterrent for passive nets); (4) ‘fishermen’s awareness and attitude regarding turtle conservation’, asked about the interviewee’s willingness to pursue responsible fishing and turtle conservation; and, finally, (5) ‘participation and cooperation’, tested their real interest in participating and cooperating in turtle conservation initiatives and research projects.

The questionnaire was designed to be completed in 15 min. Most questions were closed, which allowed collecting quantitative and factual information; some questions exploring fishermen’s opinion, in sections (3) and (5), were multiple choice questions that allowed greater freedom to their answers.

Ethics statement

All necessary permits were obtained for the field studies described. Interviewees were informed of the purpose of the study and that the data collected were confidential, and that their anonymity would be protected. The interviews were carried out only after fishermen verbally consented to participate.

Sampling methods

The study included the fishing gear types most commonly used, which were identified by a literature search (Lucchetti & Sala, 2010; Casale, 2011); those for which bycatch had been reported sporadically or not at all (i.e., purse seines) were excluded. Then, the fishermen to be interviewed were identified via opportunistic and ‘snowball’ sampling, which is widely used in sociological research (Biernacki & Waldorf, 1981; Faugier & Sargeant, 1997) and consists in the recruitment of future subjects from the acquaintances of fishermen interviewed before, overtaking the initial typical fisherman’s confidentiality. This kind of sampling permitted to reach a reasonable number of interviews. Fishermen were approached at local harbours or in public places. Since crews generally consist of several fishermen, one single fisherman per vessel was interviewed.

Data analysis

The data collected in the study related to 2014. Two types of information are required to obtain quantitative estimates and spatial data on bycatch (Moore et al., 2010): the measure of the fishing effort and the bycatch rate.

Data on the fishing effort were obtained from the EU Data Collection Framework (EU, 2008a), set up in 2000, through which Member States collect, manage, and make available a wide range of fisheries data (including biological and socio-economic data) that are needed to obtain scientific advice. Data were collected on the basis of national programmes. Disaggregated data from the DCF dataset, provided by the Italian Ministry for Agricultural, Food and Forestry Policies, were extrapolated to obtain two indexes of fishing effort: the total number of fishing vessels operating in each season in the Italian GSAs and the number of days at sea recorded in 2014.

Data on sea turtle bycatch were obtained from interviews with fishermen. The bycatch estimates, per GSA and season, were obtained by averaging (geometric mean) the number of turtles reported by fishermen for each GSA, season and gear. The geometric mean was considered to smooth the effect of the extreme values. Then, the average number of turtles caught per season, area and gear from a single boat was multiplied by the total number of vessels that actually worked with that gear in each season and area, to obtain the total number of turtles caught in Italian waters.

Data from interviews were considered as valid only if, for each combination of gear-season-GSA, at least 5 interviews were performed. Turtle bycatch was rated on a scale from 1 (low bycatch: 0–100 turtles) to 6 (very high bycatch > 1,000 turtles).

Fishermen were also asked to report the percentage of death turtles at the end of the gear retrieval (mortality rate). This value was considered to estimate the number of deaths (estimated turtle bycatch × mortality rate). The mortality rate and the estimated death of turtles obtained in the current study were then compared with those reported by Casale (2011), who made a complete review of the mortality rates for different areas of the Mediterranean Sea.

Due to the nature of data obtained by the interviews, characterized by an excess of zeros (about 75% of the entire dataset), a zero inflated model regression analysis based on negative binomial distribution (ZINB) with logit link was performed to reduce overdispersion of variance due to the zeros (Zuur et al., 2009). In ZINB analysis the zeros and the counts are analysed as two different datasets: a binomial generalized linear model (GLM) is used to model the probability of measuring a zero (called false zero, generally due to design errors, observer errors, unsuitable habitat and so on); the count process was modelled by a negative binomial GLM and as such, under certain covariate conditions, can produce zeros (in the sense that we can count zero turtles effectively and referred as true, or structural zeros). The expected mean and variance for the ZINB model are calculated as follow: ${\displaystyle E\left(Y_i\right)={\mu }_i\times \left(1-{\pi }_i\right)}$ ${\displaystyle var\left(Y_i\right)=\left(1-{\pi }_i\right)\times \left({\mu }_i+\frac{{\mu }_i^2}{k}\right)+{\mu }_i^2\times \left({\pi }_i^2+{\pi }_i\right)}$ where $E\left(Y_i\right)$ is the expected value of the response variable, μ_i is the mean of the positive count data and π_i is the probability to have false zeros. We also calculated the probability functions of ZINB to have zeros (true zeros) and negative binomial distribution for the count data as follows: ${\displaystyle f\left(y_i=0\right)={\pi }_i+\left(1-{\pi }_i\right)\times {\left(\frac{k}{{\mu }_i+k}\right)}^k}$ ${\displaystyle f\left(y_i|y_i>0\right)=\left(1-{\pi }_i\right)\times \frac{\left(y_i+k\right)!}{\left(k\right)!+\left(y_i+1\right)!}\times {\left(\frac{k}{{\mu }_i+k}\right)}^k\times {\left(1-\frac{k}{{\mu }_i+k}\right)}^y}$

where k is called the dispersion parameter.

The covariates considered for modelling data were GSA (7 levels), Season (4 levels) and Gear (3 levels). The model selection was performed following a backward selection of covariates starting from a full model with all the covariates and interactions and assessed via Akaike’s Information Criterion (AIC); the model with lowest AIC was considered the best. To assess the significance of each single factor and interactions a likelihood ratio test (LRT) based on Chi² distribution was used dropping each term in turn (Zuur et al., 2009). The best model selected consisted in the interaction between Gear and GSA (GSA × Gear) in the “count” part and the interaction between GSA and Season plus Gear as single factor (GSA × season + gear) in the “zero” part. The statistical analysis were conducted with R (v. 3.3.2; R Core Team, 2016) using the pscl package (v. 1.4.9; Zeileis, Kleiber & Jackman, 2008).

Data on stranded turtles, obtained from the Coast Guard Marine Environment Department (Reparto Ambientale Marino) database of the Italian Ministry of the Environment, were also considered as a rough index of turtle presence and abundance, and were ranked using a 6-point scale from 1 (low strandings: 0–25 turtles) to 6 (high strandings >150 turtles), to confirm the presence of hotspot areas and periods.

An interaction matrix was finally developed to find hotspot areas and periods of interaction between fishing gears and turtles for each gear. For the calculation of this matrix, fishing effort was expressed as total number of fishing days per season. The matrix was then calculated by dividing the average catch obtained by GSA, gear and season for the fishing effort expressed as total number of fishing days. The bycatch-effort interaction was also ranked from 1 (lowest risk of interaction, 0.000–0.018) to 6 (highest risk of interaction, >0.08).

The data on fishing effort, bycatch, stranded turtles and interaction matrix were plotted on separate maps using QGIS 2.8 software.

Sea turtle bycatch estimates

The fishing effort per gear type (fishing days/season/gear) calculated in the different GSAs varied greatly among seasons and gears (Table 1). For set nets it seemed to be higher in GSA 19 (Ionian Sea), given the large number of vessels operating there, especially in spring and summer. For trawling, the fishing effort was the highest in GSA 17 (Northern Adriatic Sea), where the low depth, flat seabed is ideal for towed gears, but fell in summer due to the closed fishing season. The fishing effort of longlines was consistently low, except in GSA 19 in summer.

The mean number (geometric mean) of turtles captured per GSA, fishing gear and season, obtained by the fishermen interviewed, is reported in Table 2.

Relying on interview data 52,340 capture events are estimated to occur in 2014 (Table 3). The majority of incidental catches took place in summer (>15,000 events), followed by autumn and spring (around 13,600 and 13,000 respectively), whereas a lower number were caught in winter (around 11,000) (Table 2). Catches by trawl nets mainly occurred in GSAs 17 and 18 (Adriatic Sea), where they seemed to be numerous throughout the year. Longline bycatch mainly occurred in GSAs 19 and 16 (Southern Italy), especially in summer and, to a lesser extent, in autumn. Set nets seemed to interact with turtles in most GSAs especially in spring and summer, when fishing with this gear is most active due to favourable sea and weather conditions.

The mortality rates obtained from fishermen’s interviews enabled to estimate a total of about 10,000 turtle deaths, most of them due to set nets (5,743) and trawl nets (3,082). By applying the mortality rates reported by Casale (2011) it is possible to estimate that about 21,000 turtles can die every year mainly due to set nets (around 14,000 deaths) and trawl nets (around 4,000 deaths).

The data on stranded turtles, which were especially high in summer and autumn (Fig. 2), confirmed that the Adriatic Sea is the area mostly affected by incidental catch.

The results from ZINB model are summarized in Table 4. All the factors were highly significant. The factor Season did not appear in the count part of the model indicating that it did not influence on the amount of turtle bycatch. On the other hand, Season appears as an important factor, together with the GSA and Gear, in modelling the zero distributions determining the probabilities to find false zeros. In Fig. 3, the probabilities to observe false zeros (the “zero” part of the model) are shown. The longline showed the highest probabilities that the zeros counted were false zeros in all seasons with the exceptions of summer for GSAs 9, 11 and 19. Regarding the passive nets, the probabilities of false zeros are lower than the previous although still high with only few GSAs showing less than 50% of probabilities, especially in summer and spring. On the contrary, trawl nets showed the lowest probabilities (less than 50%) to record false zeros in all seasons.

Figure 4 shows the catch predicted values per boat of the count part of the ZINB model. The highest predicted values were observed for the longline fishery especially in GSAs 9, 17 and 19. Smaller values were observed for the other two gears showing similar catching values between them, although in GSAs 17 and 18 trawl net estimates were higher.

Figure 5 reports the probabilities to catch turtles calculated by the count part of the ZINB model. What emerged was that, although, the longline predicted values were the highest, the probabilities to catch turtles by means of this gear are very low in all the GSAs (almost < 5%, except in the GSA 10). In other words, it seems that longlines have less probability to bycatch turtles but when they do that it is in relatively massive amounts. On the contrary, notwithstanding the low predicted values for the other two gears, their probabilities to accidentally catch turtles are extremely higher than longlines (between 6 and >15% for trawl nets and between 3 and 12% for passive nets).

Following these results it is clear that the major risks of turtles bycatch were associated to the trawl nets in all the GSAs (especially in the GSAs 9, 19, 18 and 17) followed by the passive nets (apart GSA 16 where the bycatch probabilities were very low) and lastly the longlines showed very low probabilities to catch turtles except in GSAs 10 and 11 where all the three gears showed almost the same probabilities.

The interaction matrix identified the gears, areas, seasons at the highest risk of bycatch (Fig. 6).

Longlines pose a risk especially in GSA 19 and, to a lesser extent, GSAs 9 and 19 in summer. Interactions with set nets cause the greatest concern in GSAs 17 and 10 in summer. Finally, the whole Adriatic Sea is an interaction hotspot for trawl nets, especially in spring and summer.

The fishermen’s perspective