Quantifying local ecological knowledge to model historical abundance of long-lived, heavily-exploited fauna

Study site

To demonstrate our methods, we used the case of the green turtle in Bahía de los Ángeles (BLA), Baja California, Mexico (28°57’6.90"N, 113°33’44.76"W), an index foraging area in the Gulf of California (Seminoff et al., 2003, 2008). We define an index foraging area as a site that (i) has aggregations of turtles in the marine environment that represent a significant proportion of the regional population, and (ii) has been monitored systematically and constantly over a prolonged period of time (>10 years). In-water scientific monitoring in this foraging area began in 1995, after population collapse (Seminoff et al., 2003, 2008). Contemporary scientific monitoring uses catch-per-unit-effort (CPUE) as a measure of abundance (Seminoff et al., 2008).

Green turtles have been a key food source for humans in the arid Baja California peninsula since the earliest phases of human occupation at least 12,000 years ago (cf. Early-Capistrán, 2014). From the late 18th century until the early 1950s, green turtle harvests were primarily subsistence-oriented. Turtles were harpooned from small, wooden canoes propelled with oars or paddles. During the 1960s, the economic and demographic growth along the U.S.-Mexico border led to an increased market for green turtle meat in Mexican border cities. BLA was a key supplier within this trade, and was able to meet demands as the introduction of outboard motors, fiberglass vessels, and set-nets increased cargo volume and catch efficiency. Additionally, improvement of transport and communication infrastructure facilitated market access (Early-Capistrán et al., 2018). The fishery collapsed in the 1970s, green turtle licenses were suspended in 1983 as populations reached dangerously low levels, and all sea turtle fishing in Mexico was banned in 1990 (Márquez, 1996; Seminoff et al., 2008).

Background research

This research is part of an on-going collaborative process in the community of BLA which began in 2012 and has included ethnographic and historiographical research related to human–ocean interaction, along with a review of scientific literature (Early-Capistrán et al., 2018). Background research helped define specific research questions, identify challenges in the study design and methods, and develop a general approach for integrating multiple forms of knowledge (Crandall et al., 2018; Early-Capistrán et al., 2018).

Historiographical research situates biological questions in a socio-historical context, providing information on a species’ past abundance which can be correlated with time-frames, social processes, and management regimes (Article S1) (Crandall et al., 2018; Sáenz-Arroyo et al., 2005). Historiographical research helped us understand human-green turtle interactions in BLA over the past three centuries, identify the early 1960s as a period when human impacts precipitated a major decline in green turtle abundance in BLA, and establish the early 1950s as a time-frame for reconstructing baseline abundance before large-scale commercial exploitation (Early-Capistrán et al., 2018).

Long-term collaboration with the community of BLA was fundamental for previously establishing the rapport and working trust necessary to conduct transdisciplinary research. Long-term engagement has also helped us acquire sensitivity to the cultural context, gain an understanding of social conditions, and gather locally-relevant information to define research questions and design (Bernard, 2011; Crandall et al., 2018). We also established a network of local collaborators, whom we define as knowledgeable community members willing to share their knowledge and expertise (Crandall et al., 2018). Due to the fact that ecological knowledge is differentially acquired by social actors, we constructed a heterogeneous network of social actors with diverse types of knowledge that, when nested together, construct the ecological knowledge around green turtle abundance (cf. Brown, 2010).

Experimental design

Reconstructing green turtle abundance through collective knowledge

Defining an approach to estimate green turtle abundance based on CPUE was a key challenge. Although CPUE is a crude measure of changes in exploited populations (López-Castro et al., 2010), we used it because (i) it is the only available metric of current abundance and (ii) CPUE is an accepted proxy for abundance for IUCN Red Listing (IUCN, 2019; O’Donnell, Pajaro & Vincent, 2010).

Adequate assessment of CPUE as a measure of abundance requires detailed understanding of the fishery and the variables that affected it (Moller et al., 2004). The skilled turtle fishers of BLA almost always targeted high-density locations (hot-spots) and aggregations, and thus maximized CPUE by optimizing fishing patterns based on empirical knowledge of environmental conditions and green turtle behavior (Early-Capistrán et al., 2018). Consequently, turtle fishers’ expertise allowed for high CPUE events over time despite declining overall abundance (hyper-stability), underscoring the need to (i) account for this non-random search behavior and (ii) understand central CPUE trends rather than exceptional catches (Article S1; Fig. S1) (Anticamara et al., 2011; Early-Capistrán, 2014; Maunder & Punt, 2004; Selgrath et al., 2018; Walters, 2003).

This scenario is challenging, as (i) interviewees’ memory of “typical” events may be less accurate than that of salient events and (ii) high variability in CPUE and changes in fishing efficiency can mask overall abundance trends (Maunder & Punt, 2004; De Damasio et al., 2015; Sáenz-Arroyo & Revollo-Fernández, 2016). Thus, we designed our methodology to calculate CPUE based on multiple sources rather than individual recollections. We also aimed to identify and account for sources of variation in CPUE that could bias proportionality with abundance, and to construct adequate proxies for variables such as spatial distribution of fishing, differences in gear types, and changes in fleet conditions (Walters, 2003; Maunder & Punt, 2004; Anticamara et al., 2011; Selgrath, Gergel & Vincent, 2018).

We approached CPUE as a component of a holistic dataset on human-environment interaction, and aimed to synthesize quantitative values on the basis of biocultural consensus, which we define as the pooling of information for evaluating shared environmental perceptions constructed by the summation of individual, community, specialist, and holistic types of knowledge. Biocultural consensus is a synergistic, interconnected set of contents and types of knowledge (c.f. Brown, 2010) in which the resulting knowledge is greater than sum of its parts. In this case, we used knowledge from all three social groups (turtle fishers, key local collaborators, and community members) as inputs for constructing biocultural consensus. Our ethnographic research was primarily focused on turtle fishers, who provided the majority of qualitative and numerical data, as well as specialized LEK related to human-turtle interaction. Key local collaborators and community members provided contextual and complementary data (Fig. 2). Biocultural consensus helped build conceptual frameworks for modeling, establish limits and assumptions, estimate model parameters, and validate model outputs (Bélisle et al., 2018).

As the primary response variable, we aimed to calculate representative values of CPUE during a specific year with the initial definition: (1) $$\mathrm{C}\mathrm{P}\mathrm{U}\mathrm{E}=\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{t}\mathrm{l}\mathrm{e}\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{u}\mathrm{g}\mathrm{h}\mathrm{t}/\mathrm{u}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{t}$$

For initial inquiry, we used the working definition of one unit effort as one night of fishing (∼12 h) with either a harpoon or a set-net (Maunder & Punt, 2004). We continually refined and updated this definition as we gained further information on fishing technology, effort, and efficiency through the iterative feedback process between qualitative data, NLR, and GLMs (Phase 2). We then standardized CPUE estimates to account for differences in gears and changes in efficiency (Phase 3). As the final result of the iterative feedback process, we obtained standardized, representative mean CPUE values for a specific year, based on biocultural consensus of green turtle captures.

Documenting LEK

M.M.E.C. and G.G.M. compiled ethnographic data in BLA over three field seasons (spring 2017, summer 2017, and spring 2018) and 57 working days. We obtained oral informed consent from all participants prior to the start of interviews, and recorded interviews in audio or video and took technical photographs when possible (Article S1; Tables S2 and S3) (International Society of Ethnobiology, 2006). We chose oral consent as it was not deemed culturally appropriate to ask participants to sign a consent document and because some participants were not comfortable with written language (International Society of Ethnobiology, 2006; Wedemeyer-Strombel et al., 2019). We conducted all interviews in Spanish—our primary language and that of the collaborators—and transcribed recorded interviews in digital format (.txt). We also compiled field journals in digital format (.txt), recording all observations in detail (Article S1).

We validated ethnographic data through triangulation among (i) participants (e.g., data were independently corroborated and verified by multiple local collaborators), (ii) sources (e.g., documents, photographs, scientific literature, etc.), and/or (iii) methods (e.g, interviews, archive research, etc.). Once processed, we member-checked data for reliability by asking local collaborators from all groups if our themes or categories were locally relevant and congruent. We also asked local collaborators to identify data gaps, and inquired if overall accounts and processes were described in a manner that was realistic and accurate (Creswell & Miller, 2000; Tengö et al., 2014). Prolonged engagement in the field allowed us to compare interview data with observations, and helped build trust so that participants were comfortable disclosing information, increasing reliability in responses (Bernard, 2011).

We identified turtle fishers using a deliberate hierarchical sampling method (Bernard, 2011), Turtle fishers are a small group of the oldest fishers in the community, between 55 and 85 years of age (N_fishers = 17). We interviewed 94% of turtle fishers, as one fisher chose not to participate. All fishers in the population and sample were men. With this target group, we continuously carried out participant observation, and conducted 17 semi-structured interviews (at least one per person), along with 27 informal interviews. Within this target population, we identified a subset of seven expert LEK holders, which we defined as turtle fishers recognized as experts by at least two peers, and whose empirical and specialized knowledge can be used as a basis for inferences and assessments about their surrounding environments and biota (cf. Bélisle et al., 2018). With the group of expert LEK holders, along with the aforementioned methods, we conducted seven in-depth interviews and one focus group discussion to gather specialized data (Tengö et al., 2014).

We identified key local collaborators (n_klc = 7) through purposive and respondent-driven sampling (Bernard, 2011). Key local collaborators were primarily older (>63: 71%) and included women (43%) and men (57%). We continuously carried out participant observation with this group, and conducted four in-depth interviews and 23 informal interviews. Topics included: local history, economy, commerce, and foodways; marine and terrestrial ethnobiology and conservation; and commercial and sport fishing, among others, which provided valuable information for situating green turtle fishing within a broader socio-ecological context (Crandall et al., 2018).

We selected local collaborators from the community at large (n_cm = 48) through a combination of cluster sampling and self-selection (Bernard, 2011). They represented ~8% of the population of BLA and included women (42%) and men (58%). Ages ranged from 18 to 93, with young (18–39: 35%), middle-aged (40–62: 37%), and older (>63: 28%) participants. While we did not inquire about income given local social taboos, local collaborators came from across all class strata with schooling varying from individuals without formal schooling to graduate degree holders. The group included both long-term residents (89%) and short-term residents (11%) such as conservation workers and government employes. This diverse group provided a broad view of perspectives and topics to complement and contextualize information from the target population of turtle fishers. With this group, we continuously carried out participant observation and conducted 72 informal interviews.

Cataloguing LEK

We processed and coded all field journals and interview transcriptions following a standardized protocol. We used footnotes to separate observations from analysis, and for cross-referencing. Cryptic indicators ensured local collaborators’ anonymity (Bernard, 2011). We used cultural material codes (Murdock et al., 2008) to categorize ethnographic data, with customized codes for topics and themes specific to this research. We indexed text entries using hashtags (#) to mark relevant topics (e.g., #fishing_gear), including ordinal codes (e.g., #max_cpue; #min_cpue) to classify information for data-binning (Article S1; see Table S4 for an example field journal entry). Along with data compiled in the 2017 and 2018 field seasons, we coded and indexed ethnographic materials collected since 2012 for integration into the qualitative database (Article S1; Tables S2 and S3). Coding allowed us to break down qualitative data into analytical variables and raw values (Strauss & Corbin, 1994). Digital files allowed for analyzing large volumes of information by facilitating topic-specific searches, generating a corroborated, systematized, and cross-referenced qualitative database (Bernard, 2011).

Synthesizing and quantifying LEK

CPUE calculation and preliminary database generation

To deal with variability, we used heuristic rules to make systematic inferences based on the knowledge of expert LEK holders (Fig. 3). This framework allowed us to calculate a central tendency based on collectively-generated knowledge and biocultural consensus rather than individual recollection, thus reducing individual cognitive bias (Article S1).

We converted captures reported by weight to number of turtles by dividing vessel capacity by mode of turtle mass (50 kg) reported by fishers and corroborated with monitoring data (Early-Capistrán et al., 2018) (Article S1). While turtle size was highly variable and likely declined in response to increasing fishing effort (Table 4), mixed juvenile/adult foraging groups with a slight juvenile bias—such as BLA, where ~56% of individuals are juveniles (Seminoff et al., 2003)—are present in green turtle foraging habitats worldwide (Seminoff et al., 2015). Thus, we consider our assumption regarding size distribution to be adequate given the nature of the data (Table 4; Article S1).

Preliminary data evaluation

We estimated CPUE and descriptor variables through an iterative process. We stored data in .csv format and carried out all analyses in R 3.4 unless otherwise specified (R Core Team, 2019). We analyzed descriptive statistics to evaluate statistical robustness by checking data distribution, evaluating normality (Shapiro–Wilk p > 0.05), and identifying outliers (±2 SD) (Zar, 2014). Each CPUE data point was linked to a summary of qualitative and numerical data for a specific collaborator, and outlying data were contextualized and evaluated (Article S1; Table S5). Over the course of the iterative process, we discarded three CPUE values from fishers who (i) had less than 1 year of experience and (ii) were very young (10–13 years of age) when they captured turtles. During interviews, these fishers recognized that they had limited recollection of events and did not have the experience necessary to provide precise data. Statistical analysis confirmed that CPUE values provided by this group were outliers (±2 SD).

To evaluate CPUE trends, we converted values for the independent variable “year” to serial form in all analyses. We used LABFit 7.2.49 to identify five preliminary models with best fit and their respective starting values. We then ran NLR (nlstools, easynls, dplyr, car, and DescTools packages; Data and Code) to choose the model that best described the data, and evaluated residuals (Table 2). We ran NLR at each round of the iterative process to (i) evaluate the general behavior and performance of the data, (ii) identify outlier effects in residual analysis, and (iii) evaluate if the process was robust to these effects (Baty et al., 2015; Ritz & Streibig, 2008). An exponential decay model consistently showed the best fit, with the form (2) $$Y\sim \alpha \cdot e^{(\beta x)}$$where Y is the response variable, CPUE; α is a constant (intercept); β is an instantaneous rate of change in the response variable (slope); and x is the independent variable “year”.

We used GLMs with a link function for Gaussian distributions to identify significant predictor variables for CPUE (nlme, dplyr, car and DescTools packages), using log-transformed values if the CPUE distribution was non-normal (Zar, 2014). We ran backward-stepping models until we obtained a model with significant effects, a high percentage of explained deviance (D²), a relatively low Akaike Information Criterion (AIC), and robust residuals (Table 2) (cf. Maunder & Punt, 2004).

We ran a total of 36 NLR and 54 GLMs on five sequential working databases. The first three databases each corresponded to one round of the methodological cycle (Fig. 1). With each round, the working databases were updated and superseded as we incorporated new data, variables, indices, analyses, and data cleaning processes (Figs. 1, 3 and 4). The last two databases consisted of the final raw database—with LEK-derived CPUE values and heterogeneous variables for unit effort—and the standardized database with mean standardized CPUE values for each year (Data and Code). By integrating these analyses into the cyclical process, we are confident that we adequately identified confounding variables and sources of variation not attributable to changes in abundance (Hilborn & Walters, 1992).

Feedback integration

We integrated model-fitting feedback by identifying which variables and indices required further information or could be improved (Fig. 4). We integrated feedback from community members during subsequent visits to the field by sharing preliminary results and model outputs with them through narrative description, and asking for collaborators’ perspectives on validity and consistency (Bélisle et al., 2018). Local collaborators also identified gaps and provided further information (Huntington, 2000; Tengö et al., 2014). We then designed new questions based on feedback and repeated these procedures with each variable (Figs. 1 and 4).

We repeated the cyclical process of data gathering, synthesis, and quantification until reaching topical saturation (similar instances were repeated and no additional data were found with which to develop new properties), thematic saturation (additional data did not produce new emerging themes), data saturation (new data repeated what was expressed in previous data) (Saunders et al., 2018), and until model fitting did not provide significant new information.

Time frames required to reach saturation are extensive, in the order of months or years. Ethnographic fieldwork generally requires a year or more, given the extensive time required to establish rapport, obtain working knowledge and understanding of the cultural context, and to be able to ask good questions and obtain good answers (Bernard, 2011). The interview process to elicit the data presented in this article represented 57 working days over three field seasons (spring 2017, summer 2017, and spring 2018). While it may seem a rather short timeframe, it must be said that two of the authors, M.M.E.C. and G.G.M., have been conducting ethnographic work in the community since the summer of 2012, making seven trips to the region with a mean duration of 27 days, conducting a total of 378 interviews to date (Tables S2 and S3), and maintaining contact and communication with community members between field seasons. Long-term continuous interaction has allowed rapport for intelligible dialog among researchers and local community members in ways that enable elicitation of trustworthy data.

Raw CPUE database analysis

The result of the methodological cycle was a final, LEK-derived CPUE database with heterogeneous variables for unit effort (raw database) (Fig. 1). We carried out descriptive statistical analysis, NLR, and GLM analysis to evaluate the data and define standardization procedures as described in “Preliminary Data Evaluation”.

CPUE database standardization

We standardized CPUE to (i) remove most of the variation not attributable to changes in abundance by accounting for variables such as gears, fleet characteristics, fishers’ experience, etc.; and (ii) generate CPUE values that could be compared over time (Hilborn & Walters, 1992; Maunder & Punt, 2004). To choose predictor variables for standardization, we ran GLMs (nlme, car, dplyr, and DescTools packages; Data and Code) with log-transformed CPUE values and a residual correlation structure based on an auto-regressive model of order 1 (AR-1) structured by the variable “year” (Zuur, 2009). We chose predictor variables for standardization using models with significant effects, high percentage of explained deviance (D²), relatively low Akaike Information Criterion (AIC), and robust residuals (Table 2) (cf. Maunder & Punt, 2004).

We generated detailed definitions of unit effort based on these analyses, in order to obtain comparable values for turtles caught in one night. While fishers generally worked from dusk to dawn, fishing times on any given night with either gear type could be variable. For modeling purposes, we simplified values to 12 h blocks which reflect the vast majority of fishing effort (Article S1).

For set-nets, we standardized unit effort to approximate ecological monitoring data (100 m net soaking for 12 h) (Koch, 2013; Seminoff et al., 2003): (3) $$C_{\mathrm{s}\mathrm{t}}=(t\cdot R)/(n_r\cdot R\cdot 12\mathrm{h})$$where C_st is a standardized, representative value of CPUE during a specific year (turtles 12 h⁻¹); t is the number of turtles caught (turtles); and n_r is the number of 100 m nets (no units). R is net length (in multiples of 100 m), simplified to short (~100 m = R) or long (~200 m = 2R) (Table 3). Soaking time is 12 h.

For harpoon captures, we assigned a skill coefficient (s, percentage of success) (Table 3) to each harpooner through social network analysis (Table 1), based on colleagues’ assessment, such that: (4) $$C_{\mathrm{s}\mathrm{t}}=t\cdot {\mathrm{s}}^{-1}\cdot 12{\mathrm{h}}^{-1}$$

The current ban on sea turtle fishing does not allow us to test for differences in susceptibility to fishing gears. Harpoons and nets were not used simultaneously by any given fisher, and both were used over a roughly equivalent number of hours per night. Thus, we considered these values to be adequately standardized given the nature of the data. For years with multiple CPUE values, we calculated the mean after standardization (Article S1; Figs. S4 and S5).

Evaluating statistical robustness

We evaluated reliability through comparison with green turtle fishery statistics for BLA (annual landings in tons, 1962–1982) (Márquez cited in Seminoff et al. (2008)). CPUE and total landings are both crude indicators of abundance, and comparative analyses have been used to assess the accuracy of LEK-derived data (De Damasio et al., 2015; Sáenz-Arroyo & Revollo-Fernández, 2016). We compared the catch reduction rate and fitted an exponential decay model (QtiPlot 0.9.9.7) as an experimental process to evaluate trends in LEK-derived CPUE and annual landings (Article S1). We then standardized both datasets to z-scores to avoid effects from differences in scales (Fig. S6) and used the Lin Concordance Correlation Coefficient (Lin CCC) to assess agreement between paired values (DescTools package; Data and Code) (Lin, 1989; Altman & Altman, 1999) (Article S1; Fig. S6).