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

Deriving robust historical population trends for long-lived species subject to human exploitation is challenging in scenarios where long-term scientific data are scarce or unavailable, as often occurs for species affected by small-scale fisheries and subsistence hunting. The importance of Local Ecological Knowledge (LEK) in data-poor scenarios is increasingly recognized in conservation, both in terms of uncovering historical trends and for engaging community stewardship of historic information. Building on previous work in marine historical ecology and local ecological knowledge, we propose a mixed socio-ecological framework to reliably document and quantify LEK to reconstruct historical population trends. Our method can be adapted by interdisciplinary teams to study various long-lived taxa with a history of human use. We demonstrate the validity of our approach by reconstructing long-term abundance data for the heavily-exploited East Pacific green turtle (Chelonia mydas) in Baja California, Mexico, which was driven to near extinction by a largely unregulated fishery from the early 1950s to the 1980s. No scientific baseline abundance data were available for this time-frame because recent biological surveys started in 1995 after all green turtle fisheries in the area were closed. To fill this data gap, we documented LEK among local fishers using ethnographic methods and obtained verified, qualitative data to understand the socio-environmental complexity of the green turtle fishery. We then established an iterative framework to synthesize and quantify LEK using generalized linear models (GLMs) and nonlinear regression (NLR) to generate a standardized, LEK-derived catch-per-unit-effort (CPUE) time-series. CPUE is an index of abundance that is compatible with contemporary scientific survey data. We confirmed the accuracy of LEK-derived CPUE estimates via comparisons with fisheries statistics available for 1962–1982. We then modeled LEK-derived abundance trends prior to 1995 using NLR. Our model established baseline abundance and described historical declines, revealing that the most critical (exponential) decline occurred between 1960 and 1980. This robust integration of LEK data with ecological science is of critical value for conservation and management, as it contributes to a holistic view of a species’ historic and contemporary conservation status.


EPISTEMOLOGY
We engaged with epistemology, ontology, phenomenology, hermeneutics, and heuristics to inform which methodologies were best suited and how they could be integrated (Moon et al., 2019). We integrated multiple ways of knowing through a common epistemology developed ad hoc through an active dialogue between knowledge systems and disciplinary approaches (Miller et al., 2008;Tengö et al., 2014). We took a phenomenological approach to the research topic, seeking to understand commonly observed phenomena (e.g.: green turtle populations observed by commercial fishers), rather than individual experience (e.g.: memories recalled by an individual fisher) (Converse, 2012).

HISTORIOGRAPHICAL RESEARCH
Historiographical research involves reviewing primary sources such as government records or historical texts, which are read critically and situated within historical processes (Brettell, 1998;Bernard, 2011;Early-Capistrán et al., 2018). We analyzed documents' internal and external validity based on hermeneutic and semiotic analysis (Denzin & Lincoln, 1994), with sensitivity to the social, political, and historical context in which they were generated and considering the impact of cultural contact, conquest, and colonialism as historical processes that can bias texts (Brettell, 1998). We identified sources of bias (observer bias, informer bias, and authorial ethnocentrism) by systematically analyzing who collected the data; how, why, under what conditions the information was produced or , PeerJ Supporting Information: Article S1 3 collected; and towards whom the texts were directed (Brettell, 1998;Bernard, 2011;Early-Capistrán et al., 2018).

Qualitative and Quantitative Data and analyses
Ethnography allows for the capture of both qualitative and quantitative data, as well as for both quantitative and qualitative analyses of data. We used both types of data and analytical approaches throughout (Table S1).

Integrating data from 2012-2013
We integrated data and materials from previous ethnographic research on the green turtle fishery in BLA in 2012-2013 (Tables S2, S3) that included participant observation, semistructured, in-depth, and informal interviews with sea turtle fishers (n=16) and community members at large (Early Capistrán, 2014;Early-Capistrán et al., 2018). It should be noted that, unfortunately, 29% of sea turtle fishers passed away between 2012 and 2017, making this previous information particularly valuable for current research. We integrated data from past field seasons -including interviews, transcriptions, and audio and video recordings (Table S3) -by applying the same processing and synthesis as data compiled in 2017-2018. This process increased the effective sample size of sea turtle fishers (n=22) and substantially increased available materials and information. , PeerJ Supporting Information: Article S1 4

Field Journal Entries
Field notes were captured continually throughout the day, and logged by blocks of time for later entry into the field journal in digital format (.txt) (Table S4). One field journal entry was captured daily. Each entry included all notes taken as well as additional information observed throughout the day. Field journal entries were detailed and extensive, ranging from 2,000-10,000 words or more. Field journal writing is a key part of ethnographic research involving two to eight hours per working day, with a working day in the field lasting 12-16 hours on average (Bernard, 2011). Field journal entries were indexed and coded for future reference and classification.
The date and study site were indicated at the beginning of the journal entry. Content is grouped in blocks of time. The approximate time of day and location, along with a general description of the activity and a cryptic indicator of the collaborator(s), were included in the heading of each block of time. Categories from the Outline of Cultural Materials (Murdock et al., 2008), a series of standardized numerical cades used to organize ethnographic data, were included at the beginning of each paragraph. We added customized codes for this research (Bernard, 2011) (e.g.: 226.5.1, Historical sea turtle fishing; 441.1, sea turtle commerce; 226.8, fishing grounds). Analysis, commentary, and cross-references were separated from observations with footnotes ([1]) at the end of each paragraph.
Analyses were cross-referenced between journal entries when applicable. Specific topics of interest were indexed using hashtags (#). See Table S4 for an abbreviated example of a field journal entry.

Numerical data capture
Numerical data was captured from field journals and interview transcriptions (see bold type in Table S4). Note that we do not refer to these values as quantitative data. These initial values were corroborated, verified, and converted into reliable, quantitative data through the processes described in Phases 2-4 of the Methods section in the main text.
We compiled summary files for each fisher, with synthesized biographical, qualitative, and numerical data from field journals, notes, and interviews, as well as references to each interview, note, and journal entry associated with the fisher (Table S5).
This allowed us to quickly access and cross-reference information. Each numerical data point used in calculations and modeling processes was linked to a summary file, and outlying data could be contextualized, evaluated, and cross-referenced.

CPUE CALCULATION
Individual memory and knowledge varied among contributors, and all fishers reported substantial variability in captures throughout the fishery. Furthermore, some fishers reported average catches (a catch in a night that was neither good nor bad) or mode catches (number of turtles caught most frequently in a single night), while others reported intervals of CPUE values or trip times to fill a vessel to capacity. Thus, we developed the framework described in Figure 3 of the main text to make systematic inferences to (i) standardize response terms and (ii) deal with data gaps and calculate CPUE by complementing information from less experienced fishers with that of experts. Specific procedures are described in detail in the following sub-sections. , PeerJ Supporting Information: Article S1 8 7Commercial dynamics The green turtle fishery followed the same catch dynamic throughout all stages: fishers would make trips of varying duration until they filled their vessel or exhausted food and water supplies, and then returned to shore to deliver the catch. Turtle fishing was carried out at night. Catches from multiple crews were gathered in pens by merchants, until a sufficient number was caught to load onto 3-4 ton trucks for transport to the market city of Ensenada, near the U.S. border, ⁓700km northwest (Early-Capistrán et al., 2018). Fishing activity was generally higher in summer (Caldwell, 1963). During some stages, crews would off-load to boats (⁓7 tons) from canoes or skiffs to allow for profitable trips to more distant fishing grounds. In such cases, we verified with fishers that their reported CPUE referred specifically to their crew and vessel (canoe or skiff with corresponding gross tonnage), separately from the total for the boat.

Calculating fishing time
We obtained minimum (1 night) and maximum (10 nights) trip duration limits from interviews. Trip duration had a maximum limit because fishers carried all supplies (food, water, fishing gear, etc.) with them. Water supply was an important limiting factor in the desert environment. During semi-structured and in-depth interviews, we asked about average trip duration, spatial distribution of fishing, and travel time to the fishing grounds used most frequently during each of the fishery stages.
Fishers reported trip time (total time from leaving port until returning with a vessel at full capacity) in number of days, which we converted to total hours. We calculated 9 fishing time using a modified application of the formula proposed by Hilborn and Walters (1992) by first calculating the number of days spent at fishing grounds: T grounds = (T total -T displacement )/24 (eqn. S1) Where T grounds (days) is time spent at fishing grounds; T total is trip time (hours), total time from leaving port until returning with a vessel at full capacity; T displacement is travel time between port and fishing grounds (hours), calculated as mean displacement time to the fishing grounds used most frequently during the fishery stage.
We converted T grounds to hours and calculated active fishing time: T fishing = (T grounds -T activities ) / 12 (eqn. S2) Where T fishing is active fishing time (hours) and T activities (hours) are non-fishing activities (catch processing, sleep, meals, repairs, etc.).
We binned time budgets into two blocks of 12 hours, based on the assumption that active fishing was carried out in ⁓12 hour blocks, with remaining time budgets allocated to non-fishing activities in ⁓12 hour blocks. We based this assumption on fishers' consistent reports that fishing activity was generally carried out from dusk until dawn, regardless of gear type. However, we must point out that for any given day, fishing durations with either gear type were variable. Harpooners would work until reaching vessel capacity and were limited by weather conditions, tides, and propulsion, among other variables, meaning that active fishing time could sometimes be less than 12 hours (e.g., on a good night when vessels were filled quickly, when changing weather did not allow for continuous fishing, etc.). In the case of nets, set times could be greater than 12 hours if turtles were scarce or if 10 weather did not allow fishers to remove nets at customary 12 hour intervals. Given the general trends and the nature of the data, we consider our 12 hour blocks to be reflective of the vast majority of fishing effort.

Calculating dates for CPUE
If fishers did not recall specific dates, we used salient events in their personal life (e.g.: marriage, birth of a child, etc.) or events in the green turtle fishery (the introduction of nets, the introduction of motors, restrictions to fisheries cooperatives or temporary bans, the total ban on turtle captures, etc.) as prompts to situate responses in a time frame within fisheries stages. If, in response to prompts, fishers provided time frames or intervals rather than specific dates, we assigned dates based on end-points within the fishery stage. For example, if they referred to early or late years within a stage, we used the first year or last year respectively (e.g.: "during my first years fishing with set-nets", "during my last years as a harpooner", etc.). If they referred to the stage in general terms we used the median year (e.g.: "when I was a harpooner"). If a fisher began or ended their career during a specific stage, the respective career dates and stage dates were used as end-points. For example, if a fisher worked from the beginning of the "Overfishing (net)" stage (1966-1972) until 1968, 1966 and 1968 were used as end-points.

Calculating number of turtles caught
Most fishers reported catches in number of turtles, but some reported total weight (kg). In these cases, we calculated the number of turtles caught by dividing gross vessel tonnage by , PeerJ Supporting Information: Article S1 11 mode turtle mass (kg). Vessel types and capacity were documented through ethnographic research, and binned as an ordinal variable (see main text, Table 3). In cases where fisher off-loaded to boats, we verified CPUE values specifically for their crews' canoe or skiff.
Fishers reported turtle sizes in kilograms, as they were paid by weight. Green turtle size distribution was highly variable and likely declined with fishing effort, as captures of large (>150kg) turtles became less frequent in later years. However, we assumed that mode green turtle size was constant across fishery stages. We based this assumption on fishers' reports of consistent mode sizes across time (Early-Capistrán et al., 2018). This is consistent with mixed juvenile/adult foraging groups with a slight juvenile bias -such as BLA, where ~56% of individuals are juveniles-found in green turtle foraging habitats worldwide (Seminoff et al., 2003(Seminoff et al., , 2015. We based our value for mode size on scientific monitoring data, corroborated with the mode weight (50kg)

Adjusting for seasonality
Expert turtle fishers noted that spatial distribution of fishing and capture rates varied greatly by season due to winter dormancy behavior. In winter, turtles were less mobile and less susceptible to gear. Fishers had to travel farther to fill their quotas, and generally had smaller catches. We then standardized questions to include references to seasonality, and to focus reports of catches primarily on summer months, when fishing was most active, to 12 reduce the bias generated by changes in fishing dynamics during winter and to account for differences in seasonality.

Spatial distribution of fishing
Spatial distribution of fishing was highly variable throughout the chronology, due to the search for hot-spots and aggregations and due to the effects of seasonality and changes in abundance. However, fishers generally covered greater distances as green turtle aggregations close to port became less frequent year-round, and as outboard motors facilitated trips of greater distance in shorter times. Thus, we used propulsion methods and trip times as proxies for spatial distribution.

RESIDUAL ANALYSIS
We incorporated residual analyses throughout model fitting processes to ensure that model assumptions were met, and that to evaluate goodness of fit and model robustness. We ensured that residuals for all models reported in the results were normally distributed, randomly distributed, independent, homoscedastic, and had zero mean (e i ~ N(0, σ 2 )). Tests and criteria used to evaluate residual auto-correlation are reported in Table 2 in the main text.
Early-Capistrán et al. (2020), PeerJ Supporting Information: Article S1 13 Evaluating residual auto-correlation in NLR We integrated a novel procedure to assess residual auto-correlation in NLR models.
Currently, recommendations for evaluating residual auto-correlation in nonlinear regression are limited to visual analysis of the residual lag plot and runs test for randomness (e.g., Ritz & Streibig, 2008;Baty et al., 2015).

COMPARATIVE ANALYSIS OF CPUE-TOTAL LANDINGS
We used QtiPlot 0.9.9.7 to fit an exponential decay model (eqn. S3) separately to both standardized, LEK-derived CPUE and fisheries statistics for BLA (annual landings in tons, 1962-1982) (Márquez cited in Seminoff et al., 2008) for an experimental evaluation of general trends in both data sets: β 0 +β 1 * exp (-x/t) (eqn. S3) Where β 0 is an offset value, β 1 is amplitude, and t is e-folding time. It must be noted that our objective was not provide a single model that best described both sets, but to evaluate if both could be described mathematically in broadly similar terms, and ascertain if we could proceed to test agreement between the two datasets.
This experimental process yielded high R 2 values for both LEK-derived CPUE (R 2 =0.848) and fisheries statistics (R 2 =0.845), although residuals for LEK-derived CPUE were non-normal at 5% significance (Shapiro-Wilk p=0.02). We did not report detailed results for this model-fitting process in the main text, as it was not the best fit for LEKderived data. However, this process suggested that both datasets share a similar tendency and can be broadly described in similar terms. Based on this analysis, we considered that the two datasets have a shared tendency, and chose the Lin CCC to further analyze and evaluate agreement evaluate agreement (Lin, 1989).