Predicting defaunation: accurately mapping bushmeat hunting pressure over large areas

Abstract

Unsustainable hunting is emptying forests of large animals around the world, but current understanding of how human foraging spreads across landscapes has been stymied by data deficiencies and cryptic hunter behaviour. Unlike other global threats to biodiversity like deforestation, climate change and overfishing, maps of wild meat hunters' movements—often based on forest accessibility—typically cover small scales and are rarely validated with real-world observations. Using camera trapping data from rainforests across Malaysian Borneo, we show that while hunter movements are strongly correlated with the accessibility of different parts of the landscape, accessibility measures are most informative when they integrate fine-scale habitat features like topography and land cover. Measures of accessibility naive to fine-scale habitat complexity, like distance to the nearest road or settlement, generate poor approximations of hunters’ movements. In comparison, accessibility as measured by high-resolution movement models based on circuit theory provides vastly better reflections of real-world foraging movements. Our results highlight that simple models incorporating fine-scale landscape heterogeneity can be powerful tools for understanding and predicting widespread threats to biodiversity.

1. Background

In the face of humanity's growing demand for natural resources, tropical forests and their fauna are suffering massive declines in biodiversity. Highly visible human threats like deforestation are often blamed for biodiversity loss, but it is becoming clear that direct exploitation—hunting for human consumption—is at least as serious a threat for many large tropical species [1,2]. This can have serious implications for the integrity and functioning of tropical forest ecosystems. Game species are often frugivorous seed dispersers, and the loss of these species can trigger cascading ecological effects including altered tree recruitment [3,4] and even diminished forest carbon sequestration [5].

For wide-ranging threats like hunting, effective conservation requires an understanding of where the threat occurs and how severely [6]. Most other global threats to biodiversity such as overfishing [7], habitat degradation [8,9] and climate change [10] have been mapped on global scales. However, spatial variation in wild meat hunting effort—the amount of energy or time spent foraging—is very poorly understood, particularly in the tropics where the risk of defaunation is most severe and where hunting affects the greatest number of species [11]. Hunters' movements have been mapped almost exclusively on small scales, rarely encompassing more than a few hunting communities [12] (but see [13]). If large-scale management and mitigation of wild meat hunting are to succeed, conservationists need scalable, generalizable and accurate models to predict where hunters are foraging even when on-the-ground data are scarce.

Wild meat hunters are typically central-place foragers, and their foraging patterns should, therefore, be distributed on the landscape according to how easily they can reach forested areas that support game [12,14,15]. Based on this principle, landscape accessibility is a commonly used estimate of hunting pressure. For example, Benítez-López et al. [16] recently developed a map of hunting pressure across the tropics. For each mapped pixel, they used the distance to the nearest access point and market as a predictor of the spatial distribution of hunting pressure. But while simple Euclidean distance measures can successfully describe coarse patterns of game depletion [15], they ignore fine-scale environmental features like topography and land cover that almost certainly influence hunters’ movement decisions while foraging [14]. As an alternative to these simple distance measures, measures of landscape accessibility that incorporate non-random movements across heterogeneous landscapes may offer more realistic and generalizable predictions of the distribution of hunting effort. Circuit-theoretic methods, for example, use algorithms designed for electrical circuits to model movement of individuals through complex environments (table 1) [17]. Despite their broad use in the fields of movement ecology, biodiversity conservation and evolution, circuit-theoretic models have yet to be used to predict of human foraging.

Table 1.

Fundamental concepts of circuit-theoretic models and their extension to dispersal modelling of human hunters’ foraging.

circuit theory concept	electrical and ecological definitions
resistance	electrical: how much the flow of electrical current is reduced as it passes through a medium ecological: the difficulty a hunter has in crossing part of the landscape, measured as the amount of time required for a foraging hunter to cross a single map pixel, assuming that they are moving at the maximum possible speed
voltage	electrical: difference in potential energy between two points in an electric circuit, causing the movement of electric charge ecological: the size of the hunter pool in each settlement
current	electrical: the rate of flow of electrons and electric charge; in a parallel circuit, this current distributes over multiple pathways according to each pathway's resistance ecological: the rate of movement of hunters as they move through the landscape; because there are many paths through the landscape (i.e. parallel pathways), the distribution of current approximates the distribution of hunter movement while foraging

Presumed relationships between accessibility and hunter movements are rarely validated with real-world observations (but see [12,14]). Lack of validation could cast doubt on the predictive ability of accessibility models, doubt that may delay the conservation planning and intervention necessary to mitigate hunting's impacts.

Here, we compare simple, commonly used measures of landscape accessibility against a novel, high-resolution accessibility model based on circuit theory and assess their ability to predict camera trap detections of hunters across tropical forests in Malaysian Borneo. We assess whether wild meat hunting effort can be described by large-scale accessibility models so as to inform management and conservation in even understudied regions.

2. Methods

(a). Camera trap observations of hunters

Hunters’ site usage was estimated from camera trap data collected across Malaysian Borneo. From 2010 to 2012, 134 unbaited camera trap stations, each containing 1–2 motion-activated cameras, were placed in seven study areas in Sabah and Sarawak (figure 1) [2]. 18–28 cameras were placed in each study area, usually ≥1 km apart. Cameras were placed primarily in lowland and lower montane tropical rainforests, some of which had been selectively logged within the last 20 years, and 111 stations were placed on trails created by wildlife. Although intended to study nonhuman mammals, particularly the Sunda clouded leopard (Neofelis diardi) and its prey, these cameras also recorded images of hunters (identified by their possession of weapons or hunting dogs). Only images captured greater than 30 min apart were included as independent human observations (cf. [18]). It is possible that multiple cameras captured single hunters' movements, but we did not consider this problematic because our primary interest was quantifying the relative local abundance of hunters that pass by or forage at each site while hunting.

Figure 1.

The distribution and number of camera traps within Malaysian Borneo (mb: Maliau Basin, mu: Gunung Mulu, ut: Ulu Trusan, up: Ulu Padas, hm: Hose Mountains, ub: Ulu Baram, pt: Pulong Tau). The overview map shows the location of each study area, inset maps magnify each study area to show the locations of individual camera stations (scale bar segments indicate a 5 km distance). (Online version in colour.)

(b). Metrics of accessibility: existing measures

We then modelled hunter site usage in response to each site's accessibility and environmental characteristics. At each site, we compared four different measures of accessibility for their predictive ability (figure 2). Three of these have been used previously to model hunting pressure: urban–rural travel time, road distance and settlement distance (i.e. distance to the nearest permanent village) [12,13,16,19]. The fourth was a novel approach using the principles of circuit theory. We chose to compare only accessibility measures that are currently globally available (i.e. urban–rural travel time [19]) or those that could be calculated from freely available geospatial data (i.e. road distance, settlement distance and circuit-theoretic accessibility).

Figure 2.

Accessibility of Malaysian Borneo to human hunters according to four alternative measures: (a) urban–rural travel time, (b) road distance, (c) settlement distance, and (d) circuit-theoretic accessibility. (Online version in colour.)

Urban–rural travel time measures how much time is required to travel via roads, waterways, railways or on foot to any point in a landscape from the nearest city with over 50 000 inhabitants (figure 2a). Road and settlement distance for each map pixel were calculated with QGIS [20]. Road distance was measured as the Euclidean distance to the nearest road—paved or unpaved—as identified in a previous study of LiDAR-derived land cover in Borneo [21] (figure 2b). Settlement distance was calculated as the Euclidean distance to the nearest settlement, as identified in state-published gazetteers [22,23] or plantation [21] (figure 2c).

(c). Metrics of accessibility: the circuit-theoretic approach

In addition to these commonly used measures, we also considered whether a more complex, fine-scale hunter movement model based on circuit theory could better describe the landscape's accessibility to hunters. Circuit theory models simulate the movement of individuals as electrons that originate from electrical ‘sources’ and move through a resistant grid to ‘grounds’ where they leave the circuit. Of all possible paths from source to ground, those with relatively low cumulative resistance experience higher current flow than high-resistance alternatives.

Assuming that hunters' travel decisions are based primarily on landscape resistance, we created a high-resolution regional map of circuit-based landscape accessibility using Gflow software [24]. Simulated hunters emerged at each village in Malaysian Borneo [22,23], travelled through a resistant landscape, and left the landscape through randomly distributed ground pixels within 12 h of the hunter's home village. Landscape resistance was calculated as the inverse of maximum travel speed, either by foot or by vehicle, in each 100 m map pixel. After simulating circuit-based diffusive movement from each village in Malaysian Borneo to its surrounding landscape, we corrected for the population size of each village to create a cumulative map of relative electrical current (figure 2d). Then, we treated relative current as a measure of the relative number of hunters expected to pass through that pixel while hunting or travelling to hunting grounds—either of which could lead to the hunting of game (full methods in the electronic supplementary material).

(d). N-mixture models of hunting activity

To account for imperfect and non-uniform detection of hunters, we modelled the relative site usage of hunters at each of the camera trap stations using hierarchical N-mixture models [25]. Hunter site usage was modelled in response to one of the four accessibility measures and unique combinations of environmental covariates, and the probability of detection was influenced by the number of active hours, the presence of a trail and monsoon season (table 2; see the electronic supplementary material for full methods). We fitted a total of 139 candidate models using the R package unmarked (see the electronic supplementary material for full model fitting results; [26,27]).

Table 2.

Covariates used to estimate hunter site usage, λ, and detection probability, p, at camera traps placed throughout Malaysian Borneo. Four alternative measures of accessibility are considered (urban–rural travel time, road distance, settlement distance and circuit-theoretic accessibility), as well as environmental covariates.

site-level λ covariate (units)	observed range (total range)	data source
urban–rural travel time (min)	303: 7168 (0: 9008)	global rural–urban mapping project's minimum urban–rural travel time [19]
road distance (km)	0.03: 6.54 (0: 7.2)	Euclidean distance between the camera station and nearest road [19,21]
settlement distance (km)	0.43: 61.25 (0: 132.7)	Euclidean distance between the camera station and nearest gazetteered village or plantation boundary [21–23]
circuit-theoretic access (log10(amps))	−2.27: 1.92 (−5.56: 3.77)	this publication
logged (within the past 10 years; 0 = no, 1 = yes)	0,1 (0, 1)	collected at camera deployment
elevation (m.a.s.l)	64: 1792 (0: 4059)	collected at camera deployment
protection status (0 = unprotected, 1 = within a protected area that allows hunting, 2 = protected from hunting)	0, 1, 2 (0, 1, 2)	protected areas identified in the previous study of landcover in Bornean Malaysia [21]
site- or sampling-event level p covariate	range	data source
trail (0 = no, 1 = yes)	0, 1	presence of a game trail; collected at camera deployment
season (0 = NW monsoon, 1 = SE monsoon)	0, 1	which monsoon season (NW monsoon or SE monsoon) during the camera detection period

We then used AICc model selection to determine the most parsimonious combination of covariates to describe hunter site usage and ensured that model ranking was not sensitive to K, pcount's integration parameter. Using the top-ranked models, we then predicted the relationship between hunter site usage and each individual covariate in that model with AICcmodavg [27,28] by setting all other covariates in the model to their mean or modal value. We chose not to use the best model to predict hunter site usage across the entire region for two reasons. First, we are primarily interested in the influence of accessibility on hunter site usage, and second, the range of observed covariate values at camera trap stations was far narrower than the values across Malaysian Borneo (table 2), rendering model-predicted maps highly uncertain.

3. Results

(a). Landscape accessibility and hunter image capture

The nearly 20 000 days of active camera trapping time resulted in 19 488 images of hunters, of which 1589 were independent. Each camera trap station—active between 1 and 578 days, with an average of 146 days—captured 0–220 independent hunter images. 76 stations captured at least one image of a hunter (average number of independent images = 20.9 overall; 35.7 in Sabah, 10.7 in Sarawak).

N-mixture model selection identified two best models of hunter site usage (AIC weights = 0.38 and 0.25; table 3). Both of these models included second-order polynomial circuit-theoretic accessibility and elevation as significant predictors of site usage while detection probability was significantly influenced by the presence of a game trail and season. The top-ranked model differed from the second only by including logging history as an insignificant predictor. For all predictors that were included in both of the top fitted models, we report model-averaged coefficient estimates.

Table 3.

The top 3 N-mixture models of hunter site usage as selected by AIC, as well as the best models that include urban–rural travel time, road distance or settlement distance for comparison. For each model, both latent hunter abundance (λ) and the probability of detecting a hunter at a given camera trap (p) are modelled as a function of environmental and/or landscape accessibility predictors (described in table 2).

model formula	ΔAICc	AICc weight	log-likelihood
λ ∼ log10 (circuit-theoretic access)2 + log10 (circuit-theoretic access) + elevation + logged; p ∼ season + trail	0.00	0.38	−3600.5
λ ∼ log10 (circuit-theoretic access)2 + log10 (circuit-theoretic access) + elevation; p ∼ season + trail	0.78	0.25	−3602.0
λ ∼ log10 (circuit-theoretic access) + elevation; p ∼ season + trail	2.52	0.11	−3604.0
…	…	…	…
λ ∼ settlement distance2 + settlement distance + elevation + logged + protection status; p ∼ season + trail	1035.4	∼0	−4115.9
…	…	…	…
λ ∼ log10 (road distance) + elevation + protection status; p ∼ season + trail	1057.0	∼0	−4127.9
…	…	…	…
λ ∼ log10 (urban–rural travel time) + elevation + protection status; p ∼ season + trail	1060.5	∼0	−4130.8

Detection probability was significantly higher for camera traps placed on game trails (α = 1.17, 95% CI = 0.84: 1.5) and during the southwest monsoon season (α = 0.65, 95% CI = 0.52: 0.78; figure 3). Hunters' local abundance was non-significantly but positively influenced by first-order circuit-theoretic accessibility (β = 0.13; 95% CI = −0.01: 0.28), but positively and significantly influenced by second-order circuit-theoretic accessibility (β = 0.12, 95% CI = 0.02: 0.21). The top models estimated that hunters were most abundant in areas of high and low landscape accessibility. Elevation had a significantly positive relationship with local hunting activity (β = 0.45, 95% CI = 0.29: 0.6). Sites that had been logged within the last 10 years had less estimated hunting activity than unlogged areas, but this difference was insignificant (β = −0.28, 95% CI = −0.60: 0.04).

Figure 3.

Model averaged predictions of (a) detection probability and (b) hunter site usage in response to significant covariates in the top-ranked N-mixture models of hunter site usage in Malaysian Borneo. Solid lines and points represent model-predicted values for each covariate; error bars represent 95% confidence bands. Grey points in (b) represent estimated hunter site usage at each site in response to environmental covariates.

The top model explained a moderate portion of the variation in observed hunter site usage (McFadden's pseudo-R² = 0.158), the vast majority of which was explained by circuit-theoretic accessibility (pseudo-R² for models only including circuit-theoretic accessibility = 0.154). In comparison, models that included only logging history, elevation and protection status fit the data far worse (pseudo-R² = 0.033). Single-covariate models ranked far below models that incorporated accessibility and environmental covariates (see the electronic supplementary material for full model rankings); however, of all covariates considered, circuit-theoretic accessibility was the highest ranked single-covariate model and was by far the most explanatory of all accessibility metrics (pseudo-R² for urban–rural travel time, settlement distance, road distance and circuit-theoretic access = 0.027, 0.030, 0.026 and 0.15, respectively). Similar to the top-ranked models including circuit-theoretic access, top models for the other (substantially less informative) accessibility measures all predicted a positive influence of accessibility, elevation and protection status on hunting activity (electronic supplementary material, figures S3–S5).

4. Discussion

Unlike other major extractive threats to global biodiversity such as overfishing and deforestation [7–10], our ability to map tropical bushmeat hunting over large spatial scales has hitherto been exceedingly poor. On small scales, the relationship between landscape accessibility and hunters' foraging movements has been well-described (e.g. [12,14]), though rarely are accessibility–hunting relationships validated with field-based observations of hunters. Even where large-scale proxies have been used to map hunting pressure [13,16], these seldom include small-scale heterogeneity and its influence on hunters’ movements. To our knowledge, this is the first extension of circuit theory-based movement models to create a regional, high-resolution map of hunters' movements and the first to validate landscape accessibility models with camera trap observations of hunters. Our fine-scaled approach to measuring landscape accessibility had a positive, non-linear relationship with hunter site usage at Bornean camera traps (table 3 and figure 3). Compared with the three other candidate accessibility measures considered—travel time to the nearest city, distance to the nearest road and distance to the nearest settlement—the circuit-theoretic accessibility measure had by far the greatest ability to predict hunter site usage. Accessibility measures that incorporate such fine-scale heterogeneity are far better correlates of hunters’ movement patterns than the coarser metrics common in the wild meat literature.

However, our findings also highlight that hunters' movement decisions are based on more than ease of travel. For example, while circuit-theoretic accessibility explained a large amount of variation in local abundance of hunters, local environmental features like elevation, logging history, protection level and the presence of game trails were also influential. Simplified movement models like the circuit-theoretic approach reduce the complexity of hunters’ decision-making by assuming that landscape resistance is the most important consideration. In reality, hunters probably decide where to forage in response to a mix of environmental and social considerations, including wildlife density, game preference, taboos and motivation for hunting [29]. We do not know how these factors may have influenced hunter behaviour in this study, but it is possible that these may explain the drop in hunting activity in intermediately accessible sites. For example, hunters may move through highly accessible sites while travelling to foraging grounds, but actively forage in remote areas that have not yet been impacted by local extirpation of game. This may also explain why logged sites—despite being more accessible—were predicted to have lower hunter abundance than unlogged sites (figure 3). Tropical forest logging and degradation are often followed by intense hunting and rapid loss of game [30], and it may be the case that Bornean hunters avoid logged forests consciously if these areas no longer support abundant wildlife. Undoubtedly, many factors beyond ease of access influence where hunters choose to hunt; nevertheless, when landscape accessibility incorporates fine-scale environmental features, it can capture a surprisingly large amount of variation in the spatial patterns of hunting effort.

Although fine-scale movement modelling of hunters may seem infeasible in understudied harvest systems, the proliferation of global geospatial data is easing the burden of complex spatial simulations. All of the geospatial data used here are available on global scales except for logging road data, which is increasingly available for much of the globe through GlobCover [31], citizen-sourced mapping projects (e.g. Global Forest Watch's Logging Roads initiative, https://loggingroads.org/) or automated processing of LANDSAT data. Circuit-theoretic modelling software catered for the ecological sciences are freely available and under active development. Barriers that once prevented widespread spatial modelling are shrinking in the face of technological innovations, opening the door to broad-scale conservation planning even in severely data-limited landscapes. However, care should be taken to create accessibility measures that reflect local conditions. If hunters are primarily driven by income, or if hunters only target a few highly valuable species, our approach to measuring accessibility is unlikely to be as informative as, for example, distance to the nearest market. Ultimately, because circuit-theoretic methods can be adapted to incorporate local drivers of hunting, landscape resistances and wild game populations, this approach can theoretically be catered to any hunting system influenced by landscape accessibility.

Describing spatial patterns of vertebrate overharvest and defaunation in tropical forests has been exceedingly difficult, especially over the regional scales at which many management decisions are made. Coarse measures of accessibility are poor predictors of hunters' foraging across space. Yet accessibility models that incorporate environmental heterogeneity at the scale of hunters’ decision-making offer a powerful and widely applicable tool to aid in the management and monitoring of wild harvest systems. Even in the most understudied tropical nations, high-resolution accessibility maps can highlight where overhunting is likely to occur, helping to mitigate the serious social, economic and ecological consequences of defaunation.

Data accessibility

Summary information of human captures at each camera trapping station are provided in the electronic supplementary material but have been stripped of specific location data; model covariates, full model rankings and analysis code are also provided.

Authors' contribution

M.C.M.D. compiled geospatial data, created mapping software, designed the methodology, conducted the statistical analysis, and drafted the manuscript. J.F.B. provided supervision, collected and managed the camera trap field data, and assisted in writing. Both authors conceived the study and gave final approval for publication.

Competing interests

We declare no competing interests.

Funding

Financial support was provided by the Natural Science and Engineering Research Council of Canada, the Canada Foundation for Innovation, Panthera, the Denver Zoo and the Universities of British Columbia and Montana.