Using Environmental Niche Modelling to Investigate Abiotic Predictors of Crocodilian Attacks on People

Using environmental niche modelling to investigate abiotic predictors of crocodilian attacks on people

G EORGE P OWELL,THOMAS M. M. VERSLUYS,JESSICA J. WILLIAMS S ONIA T IEDT and S IMON P OOLEY

Abstract Crocodilians are distributed widely through the Keywords Conservation management, crocodilian conser- tropics and subtropics, and several species pose a substan- vation, environmental niche modelling, human–crocodilian tial threat to human life. This has important implica- conflict, human–wildlife conflict, spatio-temporal modelling tions for human safety and crocodilian conservation. Un- Supplementary material for this article is available at doi.org/ derstanding the drivers of crocodilian attacks on people ./S could help minimize future attacks and inform conflict management. Crocodilian attacks follow a seasonal pattern in many regions, but there has been limited analysis of the relationship between attack occurrence and fine-scale con- Introduction temporaneous environmental conditions. We use methods from environmental niche modelling to explore the rela- anaging conflicts involving wildlife is a serious chal- tionships between attacks on people and abiotic predic- Mlenge globally, especially when the conflicts pose a tors at a daily temporal resolution for the Nile crocodile direct threat to human life (Woodroffe et al., ). Over Crocodylus niloticus in South Africa and Eswatini (formerly large regions of the tropics and subtropics crocodilians are Swaziland), and the American alligator Alligator mississip- responsible for more attacks on people per year than any piensis in Florida, USA. Our results indicate that ambient other large carnivore, with incidents reported for  coun- daily temperature is the most important abiotic temporal tries (CrocBITE, ). They also attack domestic animals predictor of attack occurrence for both species, with attack such as livestock and dogs, damage fishing nets, burrow likelihood increasing markedly when mean daily tempera- through earthen dam walls, and cause general disruption tures exceed  °C and peaking at  °C. It is likely that this to movement and livelihoods by restricting the use of relationship is explained partially by human propensity to waterways that are required for travel, fishing and house- spend time in and around water in warmer weather but hold chores (Aust et al., ; Lamarque et al., ). also by the effect of temperature on crocodilian hunting be- Successfully mitigating harmful human–crocodilian in- haviour and physiology, especially the ability to digest food. teractions has important implications not just for human We discuss the potential of our findings to contribute to the safety but also for conservation. Negative impacts on people management of crocodilians, with benefits for both human have resulted in retaliatory killings and exacerbated issues safety and conservation, and the application of environ- such as overharvesting and habitat destruction (Fukuda mental niche modelling for understanding human–wildlife et al., ), which can have wider consequences for ecosys- conflicts involving both ectotherms and endotherms. tems. As the apex predator of the aquatic environments they inhabit, and as carnivores with varied diets, crocodilians play a key role in local food webs, affecting abundance and behaviour of prey species in ways that can influence

GEORGE POWELL* Big Data Institute, Li Ka Shing Centre for Health Information the function and community structure of ecosystems (van and Discovery, Nuffield Department of Medicine, University of Oxford, Oxford, der Ploeg et al., ). They also transport nutrients from UK land to water by preying on land animals, or across ecosys- THOMAS M. M. VERSLUYS* (Corresponding author) Department of Life Sciences, tems, as in cases where they travel to feed in estuarine areas Silwood Park Campus, Imperial College London, Ascot, UK E-mail [email protected] before returning to freshwater areas. Crocodilians are also

JESSICA J. WILLIAMS ( orcid.org/0000-0002-8275-7597) Department of Genetics, ecosystem engineers, modifying their environment by dig- Evolution and Environment, Centre for Biodiversity and Environment Research, ging dens, holes and tunnels, and creating nest sites. Alli- University College London, London, UK gator holes, for instance, retain water during dry seasons SONIA TIEDT School of Public Health, Imperial College London, London, UK and serve as refuges for other aquatic species (Britton, SIMON POOLEY ( orcid.org/0000-0002-0260-6159) Department of Geography, ). Birkbeck University of London, London, UK, Department of Zoology, WildCRU, University of Oxford, Oxford, UK, and School of Life Sciences, Understanding the drivers of crocodilian attacks can help University of KwaZulu-Natal, Pietermaritzburg, South Africa minimize future incidents and inform conflict management. *Contributed equally Previous analysis of crocodilian attack patterns has focused Received  March . Revision requested  May . on identifying high-risk human demographics and beha- Accepted  June . First published online  June . viours; there has been limited analysis of potential abiotic

Oryx, 2020, 54(5), 639–647 © The Author(s), 2020. Published by Cambridge University Press on behalf of Fauna & Flora International. doi:10.1017/S0030605319000681 Downloaded from https://www.cambridge.org/core. IP address: 170.106.35.234, on 30 Sep 2021 at 08:09:15, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0030605319000681 640 G. Powell et al.

environmental predictors of attack patterns, such as tem- for human safety and conservation arising from this study, perature and rainfall (Fukuda et al., ; Brien et al., ; as well as the potential application of environmental niche Das & Jana, ). There are, however, strong theoretical modelling to the analysis of conflict with other species, both grounds for the importance of abiotic conditions because ectothermic and endothermic. temperature constrains crocodilian hunting behaviour via effects on ectotherm physiology, especially metabolic func-   tion (Hutton, ; Lance, ; Seebacher & Franklin, Methods ). Similarly, people’s propensity to expose themselves to crocodilian-inhabited waterways is probably correlated Attack and biophysical data with environmental conditions. Accordingly, despite crocodilian attacks occurring year-round in many tropical regions Information on attacks by Nile crocodiles in South Africa (e.g. India and Bangladesh), documented attack patterns are and Eswatini (–) was obtained from archival highly seasonal across Australia (Brien et al., ) and are searches including the personal archive of Tony (A.C.) correlated with mean monthly rainfall and temperature in Pooley (conservation warden responsible for monitoring South Africa, Eswatini and the southern USA (Woodward crocodile attacks in Natal Province, –, who was ac- et al., ; Pooley, ). To date, studies have only explored tive in investigations until ), attack records kept at the the contribution of abiotic factors at a coarse temporal and St Lucia Crocodile Centre, Mtubatuba, South Africa, provin- spatial resolution. In particular, they have exclusively used cial and national newspapers (print and online) in KwaZulu- average monthly or seasonal data, which do not capture Natal and South Africa, and the Times of Swaziland archive in the impact of climatic changes across days. Furthermore, Mbabane, Eswatini. Online searches used the search term these data have been used to represent regional environ- ‘crocodile’ paired with ‘attack’, ‘bite’ or ‘victim’,inEnglish mental conditions, but have typically been sourced from andAfrikaans.InformationonAmericanalligatorattacksin single weather stations, limiting the ability to capture vari- Florida (–) was provided by the Fish & Wildlife ation across space. Conservation Commission (FWC, ). Incidents involving Here, we use environmental niche modelling methods to a bite were reported to the Commission by telephone, by vic- examine the relationships between abiotic and biotic vari- tims or their families, medical caregivers, or law enforcement ables and the occurrence of crocodilian attacks on people agencies. In most instances post , the Commission’slaw at a greater spatial and temporal resolution than in previous enforcement officers conducted investigations and inter- studies. Environmental niche modelling describes methods viewed victims and witnesses. Attacks by crocodilians were for inferring the occurrence of biological phenomena in en- excluded if they were not witnessed or lacked forensic sup- vironmentally heterogeneous areas and has previously been port, if spatial or temporal data were lacking, or if they were utilized to analyse human conflict with jaguars in Brazil classified as provoked. and Mexico (Zarco-González et al., ; Carvalho et al., The attack records had different spatial resolutions: geo- ), wolves in Iran (Behdarvand et al., ), leopards graphical coordinates were available for the Nile crocodile in Pakistan (Kabir et al., ) and monkeys in India attacks (Fig. ), whereas for the American alligator attacks (Beisner et al., ). Crocodilian attacks can theoretically the Florida county in which the attack occurred was re- be thought of as an entity that can only occupy a specific en- corded (Fig. ). ArcGIS . (Esri, Redlands, USA) was vironmental niche defined by particular biophysical limits; used to ensure Nile crocodile attack coordinates fell within thus, it is possible to use environmental niche modelling to the IUCN predicted distribution of Nile crocodiles in South illustrate the likelihood of attack occurrence over time in a Africa and Eswatini (IUCN, ), and within  km of a heterogeneous environment. waterbody (Defence Mapping Agency, ). We focus on two species of crocodilian that come into We compiled a dataset of biotic and abiotic environmen- regular conflict with humans: the East African Nile crocodile tal predictors that could be spatially and temporally linked Crocodylus niloticus in South Africa and Eswatini, which is to attacks. The climatic variables were mean daily tem- estimated to be responsible for more human fatalities per perature and rainfall, and mean monthly rainfall (-year year than all other crocodilians combined (Dunham et al., means). Historical data were not available for crocodile ; Wallace et al., ), and the American alligator in and alligator population densities across the study regions, Florida Alligator mississippiensis, which is one of few in- but human population density was included as a biotic pre- digenous predators that poses a substantial threat to dictor because the majority of attacks in our study areas human life in the southern USA (Woodward et al., ). were on local residents (Woodward et al., ; Pooley, We discuss the potential of our findings to contribute to ). Year and month were included, to account for tem- the management of crocodilians, especially in subtropical poral trends. and temperate areas where climatic conditions are similar Daily rainfall and minimum and maximum temperature to those in our study regions. We also highlight the benefits data were obtained from Young et al. (). We gathered

FIG. 1 The number of American alligator Alligator mississippiensis attacks (total = ) in each county in Florida, USA, during –.

data from all weather stations within a  km perimeter of the distribution (IUCN, ) of Nile crocodiles in South Africa and Eswatini (n = ), and all stations within Florida (n = ,). To account for collinearity, mean temperature was calculated from minimum and maximum FIG. 2 The locations of Nile crocodile Crocodylus niloticus values. For the Nile crocodile dataset, missing temperature attacks (n = ) in South Africa and Eswatini during –; ’ data were linearly interpolated when the gap between re- the line indicates the westernmost limit of the Nile crocodile s distribution (IUCN, ). cords did not exceed  days, increasing coverage by .%. Average monthly rainfall was calculated as means of all days within each month. Climatic variables were linked to each The final crocodile dataset comprised  attacks re- Nile crocodile attack record from the closest NOAA station. corded in South Africa and Eswatini between September Temperature measurements were adjusted based on the  and December . The final alligator dataset com- difference between the attack elevation and station eleva- prised  attacks recorded in Florida between September tion (Fick & Hijmans, ), according to the temperature  and December . Both datasets included mean lapse rate defined by the International Civil Aviation Orga- daily temperature and rainfall, mean monthly rainfall, nization (ICAO, ). Mean climatic values were calculated human population density, month and year as variables for each county in Florida from all encompassed stations and for modelling. were linked to attack records according to date and county. Human population density in South Africa was estimated for the magisterial district encompassing each attack from Background points the closest of seven historical censuses (, , , , , , ; digitized and geo-referenced by Giraut & Attack records constitute presence only data, and there is Vacchiani-Marcuzzo, ). The annual human population no reliable measure of human–crocodilian exposure to quan- density of Eswatini was obtained from UN (). Florida tify when attacks did not occur. However, when absence county human population densities were estimated using data are not available, models can be used with background contemporaneous data from the year of the attack provided points that are randomly sampled to characterize the null by the Office of Economic and Demographic Research distribution of each variable in the model (Elith et al., (OEDR, ). In addition, the total wetland area in each ; Barbet-Massin et al., ). county, taken to be a measure of potential alligator habitat Background point sampling influences the results of (and used when selecting background points, see below), environmental niche models (Barbet-Massin et al., ; was estimated from the National Wetlands Inventory (U.S. Phillips et al., ). Therefore, we used two background Fish & Wildlife Service, ). sampling methods for each species to compare and assess

the influence of background point sampling on the results affects ROC and variable importance and therefore we used (background sampling methods A and B). Each method a bootstrapping technique in which each model was fit  was used to create datasets of , background points times with independently sampled background points, and for Nile crocodiles and American alligators that could be variable importance and ROC were measured across each sampled for model fitting. For method A we stratified back- iteration. All analyses were conducted in R .. (R Core ground point sampling according to where attacks have Team, ) and boosted regression trees were fit using been reported: , Nile crocodile attack coordinates the gbm .. package (Ridgeway, ). were randomly sampled and randomly assigned dates between  and , and , American alligator background points were randomly sampled from Florida Results counties proportional to the number of reported attacks and randomly assigned dates between  and .For Both species have a seasonal pattern of attack frequency method B we sampled background points across the species’ (Fig. ). There is a significant correlation between number predicted ranges: , Nile crocodile background points of attacks and mean monthly temperature over the study were randomly selected as coordinates that fall on water period for alligators (ω()=.,P, .) in Florida bodies within the IUCN distribution of Nile crocodiles in and crocodiles (ρ()=.,P, .) in South Africa South Africa and Eswatini, and , American alligator and Eswatini, and a significant correlation between background points were sampled from all counties in number of attacks and mean monthly rainfall for alliga- Florida (Woodward et al., ). To make this method com- tors (ρ()=.,P, .) and crocodiles (ρ()=., parable between the species, the probability of a Florida P , .). county being sampled was proportional to its total wetland The crocodile boosted regression trees had a mean cross- area (i.e. counties with more wetland area had a proportion- validation ROC value of . ± SD . using background ally higher probability of selection). Covariates were linked dataset A, and . ± SD . using background dataset B. to background points following the same process as for The alligator boosted regression trees had a mean cross- attack records. validation ROC value of . ± SD . using background dataset A, and . ± SD . using background dataset B. With background point sampling stratified by attack loca- Model fitting tions (background dataset A), mean daily temperature was the most important predictor of attack occurrence based on Spearman’s Rank correlation was calculated between all vari- model contribution for both Nile crocodiles (. ± SD ables prior to model fitting (Supplementary Table ). We first .) and American alligators (. ± SD .), and this assessed attack seasonality for both species by calculating was consistent for each model fitting iteration (Fig. ). The Spearman’s Rank correlations between the total number of distributions of model contribution overlap for daily rainfall, attacks per month, and mean monthly temperature and mean monthly rainfall, human population density, year and monthly rainfall, calculated using the background datasets. month for both species, highlighting the variability intro- We analysed attack patterns at a finer spatial and temporal duced by background sampling (Fig. ). However, on aver- resolution using boosted regression trees, which were fit to age, the second and third most important predictors of Nile express nonlinearities in the data (Elith et al., ). crocodile attack occurrence were human population density We sampled background points from the background (. ± SD .), and mean monthly rainfall (. ± .), datasets following Barbet-Massin et al. (). Boosted and of American alligator attack occurrence, year (. ± SD regression trees had an equal ratio of background points .) and human population density (. ± SD .). to attacks, and were fit with -fold cross validation, as With background points sampled across the species’ recommended by Miller (), and model fit was evaluated predicted ranges (background dataset B), mean daily tem- as the mean area under curve of the receiver operating char- perature and human population density were the strongest acteristics (ROC) plot. Boosted regression tree model para- predictors of attack occurrence. For American alligators, meters that maximized ROC were selected. Each tree had a mean boosted regression tree contribution was . ± SD learning rate of ., a tree complexity of  and a bag frac- . and . ± SD . for daily average temperature tion of .. The number of boosted trees varied for each and human population density, respectively. For Nile model but were selected to minimize holdout deviance. crocodiles, mean boosted regression tree contribution was Boosted regression tree model contribution was assessed . ± SD . and . ± SD . for mean daily tem- based on the number of times the variable was selected perature and human population density, respectively. The for splitting, weighted by the squared improvement to the relative increase in the strength of human population dens- model as a result of each split, and averaged over all trees ity as a predictor when background points are sampled (Friedman & Meulman, ). Background point sampling across the species’ predicted ranges (background dataset B)

FIG. 3 Seasonal variations in the incidence of attacks by (a) American alligators in Florida during –, and (b) Nile crocodiles in South Africa during –, with mean monthly rainfall and temperature of background points (n = ,).

compared with when sampling is stratified by attack loca-

tion (background dataset A) is probably the result of an in- FIG. 4 Boxplots of variable contribution for boosted regression crease in attack likelihood variance across space. This would trees (BRTs) for Nile crocodile and American alligator attack increase the relative importance of the covariate that is the occurrence. BRTs were fit  times with a ratio of : presence strongest spatial predictor. In our case, this is human popu- to background points sampled from two background datasets. lation density, which varies greatly across state county and Background dataset A contained background points sampled magisterial district. Human population density has a non- from attack locations over the study periods. Background dataset B contained background points sampled across the species’ linear relationship with attack occurrence across the pre- predicted ranges over the study periods. dicted ranges of both species. This is probably a result of the patchy distribution of crocodilians and people, as attacks are heavily dependent on the relative local abundance of in the model. These indicate that attack likelihood increases  both and can only occur when there is exposure between markedly when mean daily temperature exceeds °C for   the two. The lack of historical fine-scale human and croco- both species, and that it plateaus above c. °C (Fig. ). dilian population density data, or other proxies for encoun- This does not mean attacks could not occur at tempera-  ter rates between people and crocodilians, probably limits tures below °C. For example, the minimum daily average the accuracy of our models, and it is possible that the inclu- temperature linked to Nile crocodile and American alligator     sion of further covariates would affect the relative impor- attacks in our dataset were . °C and . °C, respectively. tance of predictors. Boosted regression trees indicate that the probability of Discussion attack occurrence increases for Nile crocodiles and Ameri- can alligators as temperature increases. Boosted regres- Using environmental niche modelling, we explored the sion tree partial dependence plots highlight the relation- influence of abiotic and biotic environmental variables on ship between mean daily temperature and attack likelihood crocodilian attack occurrence at a finer spatial and tempo- after accounting for the average effects of all other variables ral resolution than in previous studies. Three key findings

(Lance, ). Hossain et al. () found comparable patterns for captive saltwater crocodiles. In our study, a re- markably similar relationship was evident, including comparable minimum and maximum temperature thresholds, which offers strong support for the hypothesis that ambient temperature influences attack patterns via its effect on crocodilian physiology (Seebacher & Franklin, ;Mazzotti et al., ). A related hypothesis proposes that increased attacks at higher temperatures reflect heightened crocodilian aggression during the breeding season, mediated either by the effect of testosterone on male behaviour or female nest-guarding (Pooley et al., ). However, studies on alligators indicated that testosterone was not correlated with attack frequency (Woodward et al., ), and the limited research on hor- mones and breeding undertaken on captive Nile crocodiles suggests that heightened hormone levels and breeding occur during the late winter months, when attack occurrence is low (Kofron, ; J. Myburgh, pers. comm., ). Similarly, in the USA, limited data suggest elevated testosterone levels in the early breeding season (February– March) that decline sharply over summer (Hamlin et al., ). Moreover, for Nile crocodiles, nest-guarding in most regions occurs in the hottest months, meaning many of the larger adult females are not hunting and remain on or near their nests, reducing the likelihood of attacks by these animals (Pooley et al., ;Kofron,; Combrink et al., ). From a human behavioural perspective, attack likelihood may increase with ambient temperature as a result of people FIG. 5 Boosted regression tree (BRT) partial dependence plots spending more time in and around water in hot weather, in- showing the marginal impact of daily average temperature creasing encounters with crocodilians. In South Africa and (x-axis) on attack likelihood (y-axis) for (a) American alligators and (b) Nile crocodiles after accounting for the average effects Eswatini, nearly half of all attacks occurred on weekends of all other variables in the model. The mean (solid lines) and and holidays, suggesting that human activity patterns are standard errors (dashed lines) were calculated across all  important. In further support of this, data collected from model fitting iterations. Y-axes are on the logit scale and are our study area in southern Africa suggest that Nile crocodile centred to have zero mean over the data distribution. attacks occur most often during activities such as crossing rivers, doing domestic chores, fishing and swimming (Pooley et al., ), and similar observations have been emerged for both Nile crocodiles and American alligators. made in Zambia (Wallace et al., ). This hypothesis is Firstly, temperature was the most important abiotic temporal clearly applicable to temperature-dependent activities, predictor of attack occurrence. Secondly, attack likelihood such as swimming, but is less so for river crossing and do- increased markedly when mean daily temperature exceeded mestic chores, which occur year-round (although they may  °C. Thirdly, the probability of attacks was highest above differ in duration and method at different temperatures).  °C, where it plateaued. We discuss three non-mutually Unfortunately, because of the lack of direct data on human exclusive hypotheses that may explain these findings. activity levels for the full duration of the study periods, we From the perspective of crocodilian behaviour, attack are currently unable to quantify how changes in human be- likelihood may increase with ambient temperature for haviour contribute to attack patterns. However, this high- physiological reasons. As ectotherms, the ability of crocodil- lights an important subject for future research. ians to digest food and to hunt is dependent on ambient Variable importance and predictive accuracy (measured temperatures (Hutton, ; Emshwiller & Gleeson, ; as ROC) of the models were influenced by how background Seebacher & Franklin, ). For instance, in alligators, di- points were sampled. Predictive accuracy and the relative gestive efficiency increases significantly as body temperature importance of human population density as a predictor rises from  to  °C (Coulson & Coulson, ), and they were increased for the American alligator attack model stop eating when ambient temperatures drop below  °C when background points were sampled from across all

counties in Florida (stratified by total wetland area), as Ourfindingsalsohaveimplicationsbeyondcrocodilian opposed to stratified by counties in which attacks have conflict management. The finding that crocodilian attack occurred. This is probably a result of spatial variation in patterns appear to be influenced by inherent physiological attack likelihood between counties, which is captured by constraints that are common to all ectotherms suggests that human population density because of the close association similar methods could be applied to understand negative in- between human population density and county. This is teractions with other species. For example, snakes commonly supported by the non-linear relationship observed between bite humans, their pets and livestock, and the World Health attack likelihood and human population density across Organization () estimates that c.  million snakebites Florida. Spatial variation in attack likelihood is probably a occur annually, resulting in ,–, deaths. This result of the patchy distribution of American alligators issue has been neglected, resulting in the slow development across their predicted range and resulting variation in of mitigation strategies (Mohapatra et al., ). However, as human–crocodilian exposure rates. with crocodilians, there is both theoretical and observational Our findings have implications for crocodilian conserva- evidence that behaviour in snakes is influenced by climatic tion. In particular, improving our understanding of the cli- factors, including temperature constraints on metabolism matic conditions under which negative interactions between (Saint Girons, ; Wang et al., ). Seasonal snake- people and crocodilians are more common could inform fu- bite patterns peak in the summer months in Bangladesh ture strategies for conflict management, including educating (Rahman et al., ) and Nepal (Longkumer et al., ). local communities in high-risk areas about the relationship Our methods could also be applied to local issues that affect between temperature and attack occurrence. Our findings small communities, such as attacks by Komodo dragons on could also discourage antagonistic approaches that threat- people, pets and livestock in Indonesia. These attacks were en crocodilian welfare, such as removal, extermination and reported to have affected .% of people in a study of nest destruction (Woodward et al., ). Behavioural modi- affected areas (Ardiantiono et al., ). fication strategies are likely to be more practical for irregular In addition, the current approach of linking attacks by activities that can be scheduled during the winter, such as endothermic predators to spatial variables such as topog- dam repair (see CrocBITE, , for reports of attacks on di- raphy and urban land use (Zarco-González et al., ; vers). However, the use of crocodilian-inhabited waterways Behdarvand et al., ; Kabir et al., ; Carvalho et al., for domestic chores and livelihoods could also be modified ) could be expanded to include abiotic predictors. The by organizing essential activities during cooler periods. As heat dissipation hypothesis suggests that activity in large with all conflict mitigation strategies, this comes with the species, including dangerous predators such as the tiger caveat that attacks may still be possible at colder tempera- Panthera tigris tigris, leopard Panthera pardus saxicolor, tures because of individual variation in crocodilian behav- lion Panthera leo and hunting dog Lycaon pictus, will be iour and physiology or the risk of accidental encounters limited during hotter periods because of the difficulty (e.g. stumbling upon a nest or treading on a crocodilian). of thermoregulating (Acharya et al., ; Smith & Kok, This is especially true when temperatures are relatively low ; Creel et al., ; Ghoddousi et al., ; Robertson at dawn, dusk or during the night, as crocodilian activity et al., ). A recent study exploring injuries and fatalities often increases during these periods and humans or other caused by large mammals in Nepal found that attacks prey are more vulnerable to attack because of poor light on humans by elephants Elephas maximus, rhinoceroses conditions (CrocBITE, ). More generally, in regions with Rhinoceros unicornis and tigers were all significantly higher less seasonal temperature variation, the behavioural modifi- in the winter, potentially indicating increased aggression cation strategies described above may not be effective. For among both predators and large herbivores at colder tem- example, in India and Bangladesh, mean daily temperature peratures when vigorous activity becomes less problematic rarely drops below  °C, which is within the physiological (Acharya et al., ). As with previous work on crocodi- range at which crocodilians hunt (Das & Jana, ). lians, Acharya et al. used monthly climatic data that only of- Seasonal limitations on crocodilian attacks linked to fered a coarse temporal resolution. Our methods, therefore, temperature could be affected markedly by predicted trends could provide a template for the study of human–endo- in global warming. For instance, average temperatures in therm conflict. Africa are predicted to increase by as much as – °C by In conclusion, our findings offer strong support for the  (Ziervogel et al., ; Munzhedzi, ). This is likely hypothesis that seasonality of crocodilian attacks is deter- to facilitate more regular hunting by crocodiles, potentially mined predominantly by temperature in regions where to the point that it becomes commonplace throughout the fluctuations are large enough to have significant impacts year, as it is in equatorial Africa, India and Bangladesh. on crocodilian behaviour. We provide the first evidence of Ongoing monitoring of the influence of temperature this at a high spatial and temporal resolution. Our analysis changes on crocodilian behaviour and their effect on attack demonstrates that attacks increase predictably across a patterns will be important. given temperature range and appear to be constrained by

a biologically-relevant threshold. This has the potential to Nile crocodiles (Crocodylus niloticus) in South Africa. Behavioural inform conflict management and conservation. The strong Processes, , –.  theoretical grounding for the temperature–physiology rela- COULSON, R.A. & COULSON, T.D. ( ) Effect of temperature on the rates of digestion, amino acid absorption and assimilation in the tionship suggests that our approach could be of value in alligator. Comparative Biochemistry and Physiology Part A casting light on the dynamics of conflict between people and Physiology, , –. other species, including both ectotherms and endotherms. CREEL, S., CREEL, N.M., CREEL, A.M. & CREEL, B.M. () Hunting on a hot day: effects of temperature on interactions between African Acknowledgements We thank Allan Woodward and the Florida wild dogs and their prey. Ecology, , –. Fish & Wildlife Conservation Commission for alligator attack data. CROCBITE () The Worldwide Crocodilian Attack Database. Prior to this study, SP worked on early ideas on the seasonality of crocodile-attack.info [accessed  August ]. crocodile attacks related to biophysical variables with Joshua Potter DAS, C.S. & JANA,R.() Human–crocodile conflict in the Indian and Aiden Keane. Sundarban: an analysis of spatio-temporal incidences in relation to people’s livelihood. Oryx, , –. Author contributions Conception: SP, GP; data collection: SP; DEFENCE MAPPING AGENCY () Digital Chart of the World. environmental and spatial data collection, data preparation: GP, JJW; Defence Mapping Agency, Fairfax, USA. analyses: GP, ST; writing: TMMV, GP; revisions: all authors. DUNHAM, K.M., GHIURGHI, A., CUMBI,R.&URBANO,F.() Human–wildlife conflict in Mozambique: a national perspective, Conflicts of interest None. with emphasis on wildlife attacks on humans. Oryx, , –. ELITH, J., GRAHAM, C.H., ANDERSON, R.P., DUDÍK, M., FERRIER, S., Ethical standards This research abided by the Oryx guidelines on GUISAN,A.etal.() Novel methods improve prediction of ethical standards, and did not involve human or animal test subjects species’ distributions from occurrence data. Ecography, , –. or the collection of specimens. ELITH, J., LEATHWICK, J.R. & HASTIE,T.() A working guide to boosted regression trees. Journal of Animal Ecology, , –. EMSHWILLER,M.&GLEESON,T.() Temperature effects on References aerobic metabolism and terrestrial locomotion in American alligators. Journal of Herpetology, , –. ACHARYA, K., PAUDEL, P., NEUPANE,P.&KÖHL,M.() Human– FICK, S.E. & HIJMANS, R.J. () Worldclim : new -km spatial wildlife conflicts in Nepal: patterns of human fatalities and injuries resolution climate surfaces for global land areas. International caused by large mammals. PLOS ONE, ,e. Journal of Climatology, , –. ARDIANTIONO, A., IMANSYAH,M.&ANDAYANI,N.() FRIEDMAN, J.H. & MEULMAN, J.J. () Multiple additive regression Understanding Human–Komodo Dragon Conflict in Komodo trees with application in epidemiology. Statistics in Medicine, Village, Komodo National Park, with Suggestions for Conflict , –. Mitigation and Prevention. The st Symposium on South East Asia FUKUDA,Y.,MANOLIS,C.,SAALFELD,K.&ZUUR,A.()Deador Herpetology and Envenomation, Malang, Indonesia. alive? Factors affecting the survival of victims during attacks by saltwater AUST, P., BOYLE, B., FERGUSSON,R.&COULSON,T.() The crocodiles (Crocodylus porosus)inAustralia.PLOS ONE, ,e. impact of Nile crocodiles on rural livelihoods in northeastern FWC () Florida Alligator Bite Data. Courtesy of Dr Allan Namibia. South African Journal of Wildlife, , –. Woodward, Florida Fish & Wildlife Conservation Commission, BARBET-MASSIN, M., JIGUET, F., ALBERT, C.H. & THUILLER,W. Tallahassee, USA. () Selecting pseudo-absences for species distribution models: GHODDOUSI, A., SOOFI, M., KH HAMIDI, A., LUMETSBERGER,T., how, where and how many? Methods in Ecology and Evolution, EGLI, L., KHOROZYAN, I. et al. () Assessing the role of livestock , –. in big cat prey choice using spatiotemporal availability patterns. BEHDARVAND, N., KABOLI, M., AHMADI, M., NOURANI, E., SALMAN PLOS ONE, ,e. MAHINI,A.&ASADI AGHBOLAGHI,M.() Spatial risk model GIRAUT,F.&VACCHIANI-MARCUZZO,C.() Territories and and mitigation implications for wolf–human conflict in a highly Urbanisation in South Africa: Atlas and Geo-Historical Information modified agroecosystem in western Iran. Biological Conservation, System (DYSTURB). Institut de Recherche pour le Développement, , –. Marseille, France. BEISNER, B.A., HEAGERTY, A., SEIL, S.K., BALASUBRAMANIAM, K.N., HAMLIN, H.J., LOWERS, R.H. & GILLETTE,JR, L.J. () Seasonal ATWILL, E.R., GUPTA, B.K. et al. () Human–wildlife conflict: androgen cycles in adult male American alligators (Alligator proximate predictors of aggression between humans and rhesus mississippiensis) from a Barrier Island Population. Biology of macaques in India. American Journal of Physical Anthropology, Reproductions, , –. , –. HOSSAIN, M.S., JAMAN, M.F., AHMED, M., RAHMAN, M.M. & UDDIN, BRIEN, M.L., GIENGER, C.M., BROWNE, C.A., READ, M.A., JOYCE, M.J. M. () Food consumption of saltwater crocodile (Crocodylus &SULLIVAN,S.() Patterns of human–crocodile conflict in porosus) in a reptile farm of Bangladesh. Bangladesh Journal of Queensland: a review of historical estuarine crocodile (Crocodylus Zoology, , . porosus) management. Wildlife Research, , –. HUTTON,J.() Growth and feeding ecology of the nile crocodile BRITTON () Crocodilian Species List. crocodilian.com/cnhc/csp_ Crocodylus niloticus at Ngezi, Zimbabwe. Journal of Animal Ecology, amis.htm [accessed  May ]. , –. CARVALHO, E.A.R., ZARCO-GONZÁLEZ, M.M., MONROY-VILCHIS,O. ICAO (INTERNATIONAL CIVIL AVIATION ORGANIZATION)() &MORATO, R.G. () Modeling the risk of livestock depredation International Standard Atmosphere. ivao.aero/training/ by jaguar along the Transamazon highway, Brazil. Basic and Applied documentation/books/Student_ISA.pdf [accessed  May ]. Ecology, , –. IUCN () Crocodylus niloticus.InThe IUCN Red List of Threatened COMBRINK, X., JONATHAN, K., WARNER,C.&DOWNS, C.T. () Species :e.TA. iucnredlist.org/species// Nest-site selection, nesting behaviour and spatial ecology of female  [accessed  May ].

KABIR, M., GHODDOUSI, A., AWAN, M.S. & AWAN, M.N. () RIDGEWAY,G.() gbm: Generalized Boosted Regression Models: Assessment of human–leopard conflict in Machiara National Park, R Package Version .-.. CRAN.R-project.org/package=gbm Azad Jammu and Kashmir, Pakistan. European Journal of Wildlife [accessed  March ]. Research , –. ROBERTSON, J.A.D., ROODBOL, M., BOWLES, M.D., DURES, S.G. & KOFRON, C.P. () The reproductive cycle of the Nile crocodile ROWCLIFFE, J.M. () Environmental predictors of livestock (Crocodylus nilotkus). Journal of Zoology, , –. predation: a lion’s tale. Oryx, published online  June . LAMARQUE, F., ANDERSON, J., FERGUSSON, R.A., LAGRANGE, M., SAINT GIRONS,H.() Thermoregulation in reptiles with special OSEI-OWUSU,Y.&BAKKER,L.() Human–wildlife Conflict reference to the tuatara and its ecophysiology. Tuatara, , –. in Africa: Causes, Consequences and Management Strategies. SEEBACHER,F.&FRANKLIN, C.E. () Physiological mechanisms of FAO, Rome, Italy. thermoregulation in reptiles: a review. Journal of Comparative LANCE, V.A. () Alligator physiology and life history: the Physiology B: Biochemical, Systemic, and Environmental Physiology, importance of temperature. Experimental Gerontology, , –. , –. LONGKUMER, T., ARMSTRONG, L.J. & FINNY,P.() Outcome SMITH,Y.&KOK,O.B.() A suggested thermoneutral zone for determinants of snakebites in North Bihar, India: a prospective African lions (Panthera leo Linnaeus, ) in the Southwestern hospital-based study. Journal of Venom Research, , –. Kalahari, Namibia. Pakistan Journal of Biological Sciences, , –. MAZZOTTI, F.J., CHERKISS, M.S., PARRY, M., BEAUCHAMP, J., UN (UNITED NATIONS)() World Population Prospects: The  ROCHFORD, M., SMITH, B. et al. () Large reptiles and cold Revision. un.org/development/desa/publications/world-population- temperatures: do extreme cold spells set distributional limits for prospects-the--revision.html [accessed  December ]. tropical reptiles in Florida? Ecosphere, ,e. U.S. FISH &WILDLIFE SERVICE () National Wetlands MILLER, J.R.B. () Mapping attack hotspots to mitigate human– Inventory—Version —Surface Waters and Wetlands Inventory. carnivore conflict: approaches and applications of spatial predation U.S. Fish & Wildlife Service, Washington, DC, USA. risk modelling. Biodiversity Conservation, , –. VAN DER PLOEG, J., CAUILLAN-CUREG, M., VAN WEERD,M.& MOHAPATRA, B., WARRELL, D.A., SURAWEERA, W., BHATIA,P., PERSOON,G.() ‘Why must we protect crocodiles?’ DHINGRA, N., JOTKAR, R.M. et al. () Snakebite mortality in Explaining the value of the Philippine crocodile to rural India: a nationally representative mortality survey. PLOS Neglected communities. Journal of Integrative Environmental Sciences, Tropical Diseases, ,e. , –. MUNZHEDZI,S.() Climate Trends and Scenarios, Climate and WALLACE, K.M., LESLIE,A.&COULSON,T.() Living with Impacts Factsheet Series, Factsheet  of .DepartmentofEnvironmental predators: a focus on the issues of human–crocodile conflict within Affairs, Climate Change Branch, Pretoria, South Africa. sanbi.org/ the lower Zambezi valley. Wildlife Research, , –. wp-content/uploads///ltas-factsheetclimate-trends-and- WANG, T., ZAAR,M.,ARVEDSEN, S., VEDEL-SMITH,C.& scenarios.pdf [accessed  February ]. OVERGAARD,J.() Effects of temperature on the metabolic OEDR (OFFICE OF ECONOMIC &DEMOGRAPHIC RESEARCH)() response to feeding in Python molurus. Comparative Biochemistry & Population & Demographics. edr.state.fl.us/content/population- Physiology. Part A: Molecular and Integrative Physiology, , –. demographics/index.cfm [accessed  April ]. WOODROFFE, R., THIRGOOD,S.&RABINOWITZ, A. (eds) () PHILLIPS, S.J., DUDÍK, M., ELITH, J., GRAHAM, C.H., LEHMANN, A., People and Wildlife, Conflict or Co-Existence? (Conservation LEATHWICK,J.&FERRIER,S.() Sample selection bias and Biology). Cambridge University Press, Cambridge, UK. presence-only distribution models: implications for background WOODWARD, A.R., LEONE, E.H., DUTTON, H.J., HORD,L.&WALLER, and pseudo-absence data. Ecological Applications, , –. J.E. () Human alligator conflict in Florida, USA. In Crocodiles. POOLEY, A.C., HINES,T.&SHIELD,J.() Attacks on humans. In Proceedings of the rd Working Meeting of the IUCN-SSC Crocodile Crocodiles and Alligators (ed. C.A. Ross), nd edition, pp. –. Specialist Group, –. IUCN, Gland, Switzerland. Merehurst, London, UK. WHO (WORLD HEALTH ORGANIZATION)() Snakebite POOLEY,S.() Using predator attack data to save lives, human and envenoming. who.int/snakebites/disease/en [accessed  August ]. crocodilian. Oryx, , –. YOUNG, A.H., KNAPP, K.R., INAMDAR, A., ROSSOW, W.B. & HANKINS, POOLEY, S., BOTHA, H., COMBRINK,X.&POWELL,G.() W. () The International Satellite Cloud Climatology Project Synthesising Nile crocodile attack data and historical context to H-Series climate data record product. Earth System Science Data, inform mitigation efforts in South Africa and eSwatini (Swaziland). , –. Oryx, published online  July . ZARCO-GONZÁLEZ, M.M., MONROY-VILCHIS,O.&ALANÍZ,J.() RCORE TEAM () The R Project for Statistical Computing. r-project. Spatial model of livestock predation by jaguar and puma in Mexico: org [accessed January ]. conservation planning. Biological Conservation, , –. RAHMAN,R,FAIZ, M.A., SELIM, S., RAHMAN, B., BASHER, A., JONES, ZIERVOGEL, G., NEW, M., ARCHER VAN GARDEREN, E., MIDGLEY, G., A. et al. () Annual incidence of snake bite in rural Bangladesh. TAYLOR, A., HAMANN, S. et al. () Climate change impacts and PLOS Neglected Tropical Diseases, ,e. adaptation in South Africa. WIREs Climate Change, , –.