Introduction
The dhole or Asiatic wild dog Cuon alpinus is a large (10–20 kg), wide-ranging carnivore facing global extinction. In the past, dholes occurred in large areas of alpine, temperate, tropical and subtropical forests across most of Asia (Kamler et al., Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015) but they are now confined to just 25% of their historical range, mostly within protected areas (Wolf & Ripple, Reference Wolf and Ripple2017). Existing populations are small, isolated, and often exhibit severe local population fluctuations (Kamler et al., Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015; Li et al., Reference Li, McShea, Wang, Gu, Zhang, Zhang and Shen2020). The current global population is estimated to be 1,000–2,200 adults, with further population declines projected as a result of continuing habitat loss and fragmentation, persecution, prey depletion, interspecific competition and disease (Davidar & Fox, Reference Davidar, Fox and Fox1975; Gopi et al., Reference Gopi, Habib, Lyngdoh and Selvan2012; Kamler et al., Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015; Srivathsa et al., Reference Srivathsa, Karanth, Kumar and Oli2019). These threats are expected to increase in severity with human population growth, and concrete conservation action is needed to protect the species from global extinction (Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022).
Large carnivores such as dholes are ecologically important and often act as umbrella and flagship species for conservation (Gittleman et al., Reference Gittleman, Funk, MacDonald and Wayne2001; Dalerum et al., Reference Dalerum, Somers, Kunkel and Cameron2008; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021). However, their carnivorous diet and need for large areas of suitable habitat frequently bring them into conflict with people (Woodroffe, Reference Woodroffe2000; Madden, Reference Madden2004; Chapron et al., Reference Chapron, Kaczensky, Linnell, von Arx and Huber2014). Although coexistence is possible, legal and illegal persecution sometimes happens, with associated cultural and socio-economic repercussions (Woodroffe, Reference Woodroffe2000; Treves & Karanth, Reference Treves and Karanth2003; van Eeden et al., Reference van Eeden, Eklund, Miller, López-Bao and Chapron2018; Dalerum, Reference Dalerum2021).
Carnivore conservation is a complex and resource-intensive issue where competing factors have to be prioritized (Macdonald & Sillero-Zubiri, Reference Macdonald, Sillero-Zubiri, MacDonald and Sillero-Zubiri2004; Madden, Reference Madden2004; Leader-Williams et al., Reference Leader-Williams, Adams and Smith2010). Spatial prioritization should be based on a comprehensive knowledge of the current and potential distribution of the species of conservation concern (e.g. Eriksson & Dalerum, Reference Eriksson and Dalerum2018). Environmental niche models are particularly useful tools that use ecological information to link occurrence and environmental data to understand and predict species distributions (Elith & Franklin, Reference Elith, Franklin and Levin2013; Zhu et al., Reference Zhu, Liu, Bu and Gao2013). They are used widely in ecology, evolutionary biology and environmental management to investigate a broad range of issues including biological invasions, the effects of climate change and spatial disease transmission (Zhu et al., Reference Zhu, Liu, Bu and Gao2013).
The MaxEnt algorithm is a robust method of predicting the potential geographic distribution of a species (Phillips et al., Reference Phillips, Anderson and Schapire2006, Reference Phillips, Anderson, Dudík, Shapire and Blair2017). It relies on maximum entropy to relate species occurrence data to a set of environmental predictors (Elith et al., Reference Elith, Graham, Anderson, Dudík and Ferrier2006), and belongs to a class of environmental niche models that require occurrence data only (Elith et al., Reference Elith, Phillips, Hastie, Dudík, Chee and Yates2011). Therefore, inherent issues with logistic models based on uncertain pseudo-absences are largely removed (Ward et al., Reference Ward, Hastie, Barry, Elith and Leathwick2009). Despite the rapid development of new algorithms for occurrence-only models, the MaxEnt algorithm is still among the best performing in terms of predictive accuracy, and its output is closely correlated with empirical data (Valavi et al., Reference Valavi, Elith, José, Lahoz-Monfort and Guillera-Arroita2021). Furthermore, it maintains high accuracy even with a relatively low number of occurrence records (Wisz et al., Reference Wisz, Hijmans, Li, Peterson, Graham and Guisan2008). However, as with other machine learning algorithms (Scowen et al., Reference Scowen, Athanasiadis, Bullock, Eigenbrod and Willcock2021), it tends to favour a level of complexity that renders it less useful for a mechanistic understanding of how specific environmental characteristics influence the potential for certain areas to be suitable habitat for the target species (many published MaxEnt models have well over 100 parameters).
We applied the MaxEnt algorithm to dhole occurrence data to create a map of potential range and to estimate the relative suitability of these areas (Kao et al., Reference Kao, Songsasen, Ferraz and Traylor-Holzer2020). We used a coarse-scale model to delineate the potential range and a finer-scale model to evaluate the relative probability of dhole occurrence within these areas. Previous distribution models on dholes are limited to regional or local scales (Nurvianto et al., Reference Nurvianto, Imron and Herzog2015; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Havmøller et al., Reference Havmøller, Havmøller, Nawangsari, Pratiwi, Møller and Træholt2022; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022). Our objective was to aid spatial planning and prioritization for dhole conservation across large parts of the global range, including areas not currently occupied (Guillera-Arroita et al., Reference Guillera-Arroita, Lahoz-Monfort, Elith, Gordon, Kujala and Lentini2015). Specifically, we aimed to (1) identify the spatial distribution of potential dhole range in 12 countries within the species’ known range and (2) quantify spatial variation in its relative probability of occurrence. This information is a prerequisite for effective dhole conservation management planning.
Study area
We included 12 countries in our study, which we grouped into three subcontinents based on McColl (Reference McColl2005): China (including the mainland of the People’s Republic of China, hereafter referred to as ‘mainland China’), the Indian subcontinent (including Nepal, Bhutan, Bangladesh and India), and Southeast Asia (including Myanmar, Lao People’s Democratic Republic (Lao PDR), Viet Nam, Thailand, Cambodia, Malaysia and Indonesia). Detailed descriptions of the environmental and socio-economic characteristics of these regions are available in Supplementary Material 1.
Environmental layers used to model range suitability for the dhole Cuon alpinus, as well as whether or not each variable was included in a coarse- (8 × 8 km) and a fine-scale (2 × 2 km) MaxEnt model. Only variables with a correlation of 0.8 or less with any other variable were included in each model.
1 Biomes: tropical & subtropical moist broadleaf forests; tropical & subtropical dry broadleaf forests; temperate broadleaf & mixed forests; tropical & subtropical coniferous forests; temperate conifer forests; boreal forests/taiga; tropical & subtropical grasslands, savannahs and shrublands; temperate grasslands, savannahs and shrublands; flooded grasslands & savannahs; montane grasslands & shrublands; deserts & xeric shrublands; mangroves & snow.
2 Land-cover classes: cultivated terrestrial areas and managed lands; woody trees; herbs; shrubs; natural and semi-natural aquatic vegetation; artificial surfaces; bare areas.
3 Soil categories: soils with clay-enriched subsoils; soils with little or no profile differentiation, pronounced accumulation of organic matter in the mineral top soil; soils distinguished by Fe/Al chemistry; soils with thick organic layers; soils with limitations to root growth; soils formed from the arid climate; shallow soils rich in humus formed from carbonates; soils with depth surface.
4 Protected area classes: protected; not protected.
Methods
Environmental variables and spatial scale
We selected 24 environmental variables known to influence the distribution of large, wide-ranging carnivores (e.g. Swanepoel et al., Reference Swanepoel, Lindsey, Somers, van Hoven and Dalerum2013; Eriksson & Dalerum, Reference Eriksson and Dalerum2018), many of which have previously been used to model dhole distribution over local and regional scales (Nurvianto et al., Reference Nurvianto, Imron and Herzog2015; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Havmøller et al., Reference Havmøller, Havmøller, Nawangsari, Pratiwi, Møller and Træholt2022; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022). They are associated with climate, ecology, geophysical factors and human impact. Of these, we retained 20 uncorrelated variables (R < 0.8) for the coarse-scale model and 19 for the fine-scale model (Table 1).
Species distribution models, including ones fitted using the MaxEnt algorithm, are sensitive to grain sizes, i.e. the spatial scale at which environmental characteristics are linked to species observations (Gottschalk et al., Reference Gottschalk, Aue, Hotes and Ekschmitt2011; Song et al., Reference Song, Kim, Lee, Lee and Jeon2013). We defined both coarse- and fine-scale grain sizes based on biologically meaningful information (Zarzo-Arias et al., Reference Zarzo-Arias, Penteriani, Delgado, Peón Torre, Garcia-Gonzalez and Mateo-Sánchez2019). We set the coarse-scale grain size to 8 × 8 km (64 km2), which approximates to the mean home range size reported for dholes (53.4 km2; Acharya et al., Reference Acharya, Johnsingh and Sankar2010; Jenks et al., Reference Jenks, Songsasen and Leimgruber2012; Srivathsa et al., Reference Srivathsa, Narayanarao and Karanth2017). We set the fine-scale grain size to 2 × 2 km (4 km2), which corresponds to the estimated daily movement of dholes (2.2 km; Grassman et al., Reference Grassman, Tewes, Silvy and Kreetiyutanont2005) and similar species such as the Eurasian wolf Canis lupus lupus (2.5 km; Kusak et al., Reference Kusak, Skrbinšek and Huber2005). We specified the coarse-scale model area as the entire study region, but excluded grid cells that were largely aquatic (i.e. where land comprised less than 50% of the area). We also excluded all islands smaller than 25,000 km2 because we regarded these areas as too small to hold viable dhole populations. Such small islands could act as demographic sinks and would thus not be relevant from a conservation perspective. The final coarse-scale model contained 240,970 cells of 8 × 8 km. We specified the fine-scale model area as those cells identified as potential dhole range in the coarse-scale model, resulting in 390,976 cells of 2 × 2 km. We rescaled all environmental variables to the two grain sizes using QGIS 3.26 (QGIS Development Team, 2023) and functions provided by raster 3.5-15 (Hijmans, Reference Hijmans2022) for the statistical environment R 4.2.1 (R Core Team, 2023).
Distribution of potential dhole range and the relative probability of dhole Cuon alpinus occurrence for 12 countries across the Indian subcontinent, Southeast Asia and mainland China. We estimated the distribution of the potential range across the study area from a binary classification of the output from a MaxEnt model with 8 × 8 km resolution, and the relative probability of occurrence as the complementary log–log transformation of the output from a MaxEnt model with 2 × 2 km resolution.
Dhole occurrence data and spatial filtering
We compiled a dataset of 1,604 geographical locations of dholes observed during 1996–2018 (Supplementary Material 2; Supplementary Table 2; Supplementary Fig. 1a). Data were provided by participants in a workshop co-organized by the dhole working group of the IUCN Species Survival Commission (SSC) Canid Specialist Group, the IUCN SSC Conservation Planning Specialist Group, Smithsonian Conservation Biology Institute, Kasetsart University and the Khao Yai National Park in Thailand in 2019 (Kao et al., Reference Kao, Songsasen, Ferraz and Traylor-Holzer2020).
Spatial filtering is a powerful method of reducing sampling bias to improve the performance of environmental niche models (Boria et al., Reference Boria, Olson, Goodman and Anderson2014). We filtered our raw occurrence data in two stages for each spatial scale, using an algorithm based on finding the maximum number of observations while respecting a minimum nearest-neighbour distance, implemented in R spThin 0.2.0 (Aiello-Lammens et al., Reference Aiello-Lammens, Boria, Radosavljevic, Vilela and Anderson2015). Firstly, we restricted the dataset to one observation per cell, which reduced the number of dhole observations from 1,604 to 567 cells for the coarse-scale model and to 1,011 cells for the fine-scale model. Secondly, we only included one record per 3 × 3 cell neighbourhood at the coarse scale and one record per 6 × 6 cell neighbourhood at the fine scale (i.e. if there were multiple records in such a neighbourhood, they were re-presented as a single data point in the centre of that neigh-bourhood). Therefore, the minimum nearest-neighbour distance was 12 km. The final dataset comprised 299 cells at the coarse scale (Supplementary Fig. 1b) and 291 cells at the fine scale (Supplementary Fig. 1c).
Environmental niche modelling
We ran the Java version of MaxEnt 3.4.4 (Phillips et al., Reference Phillips, Anderson, Dudík, Shapire and Blair2017), implemented in R using the packages dismo 1.3–3 (Hijmans et al., Reference Hijmans, Phillips, Leathwick and Elith2021) and ENMeval v2.0.3 (Kass et al., Reference Kass, Muscarella, Galante, Bohl, Pinilla-Buitrago and Boria2021). MaxEnt implements a maximum entropy approach to the presence-only class of environmental niche models by associating species occurrence with environmental characteristics using linear, quadratic, product, threshold and hinge features (Phillips et al., Reference Phillips, Anderson and Schapire2006). This parameterization allows for the modelling of potentially complex relationships among environmental characteristics (Elith et al., Reference Elith, Phillips, Hastie, Dudík, Chee and Yates2011). Although machine learning algorithms such as MaxEnt generally favour more complex model solutions than likelihood-based algorithms, over-fitting can still be problematic (Warren & Seifert, Reference Warren and Seifert2011). The MaxEnt software controls for over-fitting by using a regularization parameter that penalizes variables with low contribution to the model. As a MaxEnt model with any given data can have a large number of alternative parameterizations and regularization values, identification of the most parsimonious model and appropriate model tuning is important (Merow et al., Reference Merow, Smith and Silander2013).
We created a set of 310 models including combinations of all five types of feature (i.e. linear, quadratic, product, threshold and hinge features), each sequentially run over a set of regularization multipliers ranging from 0.1 to 10 for each spatial scale. We then identified the most parsimonious combination of feature types and regularization values using the Akaike information criterion corrected for small sample sizes (AICc; Akaike, Reference Akaike1974). We calculated the AICc values from raw model output where the sums of the log transformed raw values were treated as equivalent to model likelihood (Warren & Seifert, Reference Warren and Seifert2011). Following Burnham & Anderson (Reference Burnham and Anderson2002), we regarded models within two AICc units of each other as having equivalent empirical support. We evaluated model performance using the value of the area under the receiver operating characteristic curve (AUC; Fielding & Bell, Reference Fielding and Bell1997) as well as three model performance metrics based on cross-validation using a checkerboard method to separate our occurrence data into training and testing sets (Kass et al., Reference Kass, Muscarella, Galante, Bohl, Pinilla-Buitrago and Boria2021): AUCtest, which describes the ability of testing locations to distinguish between background and presence locations, AUCdiff, which describes the difference in the ability to distinguish between presence and background locations between training and test data (Warren & Seifert, Reference Warren and Seifert2011), and ORMTP, which is the proportion of test locations with a value below the lowest value of training locations (minimum training presence omission rate; Kass et al., Reference Kass, Muscarella, Galante, Bohl, Pinilla-Buitrago and Boria2021). AUC values from 0.7 to 1.0 generally suggest that the model has adequate predictive ability (Araújo et al., Reference Araújo, Thuiller, Williams and Reginster2005), whereas AUCdiff and ORMTP values substantially above zero indicate over-fitting.
Binary classification of potential range
We used the complementary log–log (cloglog) transformation of the raw MaxEnt values, which is bounded between 0 and 1, as the basis for summarizing the results (Phillips et al., Reference Phillips, Anderson, Dudík, Shapire and Blair2017). To outline the potential dhole range, we converted the cloglog output from the coarse-scale model into a binary layer using the minimum cloglog score of any cell with dhole presence, after the presence cells with the lowest 10% of cloglog scores had been omitted. This corresponded to a cloglog score of 0.24 and we classified cells at or above this threshold as potential dhole range. The outline of these areas was used as the model region for the fine-scale modelling. We evaluated the relative probability of dhole occurrence as equivalent to the cloglog values derived from the fine-scale model (Phillips et al., Reference Phillips, Anderson, Dudík, Shapire and Blair2017).
Estimation of variable contributions
We used three methods to evaluate the relative contribution of each environmental variable to the model at each spatial scale. Firstly, we used a heuristic method that estimates the percentage contribution of each variable to the MaxEnt solution as the proportional contribution to the model training gain for every iteration of the model-fitting process (Phillips et al., Reference Phillips, Anderson and Schapire2006). Secondly, we calculated the regularized training gain for each variable when used by itself, indicating how useful each variable was for the model solution. Thirdly, we used a jackknife procedure to evaluate how much regularized training gain was lost when each variable was omitted compared to when all variables were included in the model, indicating how much unique information was contributed by each variable.
Results of our analysis of the dhole’s potential range and relative probability of occurrence for 12 countries across three regions: (a) proportion of potential dhole range per region and country, (b) proportion of land area within each region and country identified as potential dhole range, (c) mean ± SD relative probability of occurrence across the regions and counties. We estimated the potential dhole range across the study area from a binary classification of the output from a MaxEnt model with 8 × 8 km resolution, and the relative probability of occurrence as the complementary log–log transformation of the output from a MaxEnt model with 2 × 2 km resolution.
Results
Model selection and model performance
The optimal coarse-scale model included linear, product and threshold features introduced through 97 parameters, and the optimal fine-scale model included linear and threshold features introduced through 87 parameters. Both models had a regularization multiplier of 1.5. The models were 13.49 (coarse-scale) and 5.16 (fine-scale) AICc units above the model with the second lowest AICc scores (Supplementary Table 2). Models at both scales showed high predictive accuracy, with AUC scores of 0.96 for the coarse-scale model (Supplementary Fig. 2a) and 0.82 for the fine-scale model (Supplementary Fig. 2b), and high mean AUC values based on the withheld testing data (coarse-scale model: AUCtest = 0.93; fine-scale model: AUCtest = 0.75). There were no indications of over-fitting for either model (low differences between the training and testing data sets in respective AUC scores; coarse-scale model: AUCdiff = 0.03; fine-scale model: AUCdiff = 0.07), as well as minimum training presence omission rates close to zero for both models (ORMTP = 0.03 for both the coarse- and the fine-scale model; Supplementary Table 2).
Distribution of potential dhole range and relative probability of dhole occurrence
We identified potential dhole range in three regions: along the west coast of India, in central east India, and across the foothills of the Himalaya and continuing south through Southeast Asia (Fig. 1). The largest area was in Southeast Asia (56% of the total potential dhole range identified) with a further 33% in India (Fig. 2a). We identified 80% of Bhutan as potential dhole range, the highest proportion of any country, and ≥ 30% of land as potential dhole range in all countries in Southeast Asia (Fig. 2b). The highest mean relative probability of dhole occurrence was in Bhutan, Thailand, Cambodia and Malaysia (Fig. 2c), and the relative probability of dhole occurrence was on average higher in Southeast Asia (0.38 ± SD 0.24) than on the Indian subcontinent (0.36 ± SD 0.23) or in mainland China (0.36 ± SD 0.17).
Per cent contribution of selected environmental variables to MaxEnt models of potential dhole range: (a) coarse scale (8 × 8 km), (b) fine scale (2 × 2 km). The contribution was based on a heuristic method that estimates the proportional contribution of each variable to the model training gain for every iteration of the model-fitting process. NDVI, normalized difference vegetation index.
Jackknife tests of contributions of selected environmental variables to MaxEnt models of potential dhole range: (a) coarse scale (8 × 8 km), (b) fine scale (2 × 2 km). Each graph shows the regularized gain when a variable is used on its own (black bars) as well as the loss in regularized gain when it is removed from the full model (grey bars). NDVI, normalized difference vegetation index.
Contributions made by environmental variables
Land protection status (coarse-scale 37%; fine-scale 59%) and temperature seasonality (coarse-scale 26%; fine-scale 13%) contributed most to the models at both spatial scales, with land protection status contributing substantially more to the fine-scale model (Fig. 3). Land protection was positively associated with dhole range suitability for both models (Supplementary Figs 3 & 4), whereas temperature seasonality showed a non-monotonic relationship with dhole range suitability in the coarse-scale model (Supplementary Fig. 3) and a bimodal relationship in the fine-scale model (Supplementary Fig. 4). Other important variables were tree cover (12%), elevation (6%), density of medium-sized livestock (4%) and annual mean temperature (3%) for the coarse-scale model (Fig. 3a), and human population density (5%), annual precipitation (5%), precipitation of the wettest month (3%) and tree cover (3%) for the fine-scale model (Fig. 4). Overall, land protection status was the most informative variable individually and carried the most unique information when combined with all other variables (Fig. 4a,b). Temperature seasonality and tree cover were important individually and contributed high levels of unique information to the coarse-scale model and, likewise, temperature seasonality, annual precipitation and livestock density contributed to the fine-scale model. Marginal response curves showing how the predicted probability of dhole presence changes as each environmental parameter is varied while keeping all other predictors constant are provided in Supplementary Figs 3 and 4.
Discussion
Most areas identified as potential dhole range were located in three major regions; one along the west coast of India, a second in central India, and a third across the foothills of the Himalayas and continuing through Southeast Asia. These regions largely coincide with those identified in earlier studies (Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022). However, these three regions are not directly connected, and dhole habitat is heavily fragmented particularly in the central Indian and the eastern regions. Hence, it is important to identify and secure dispersal corridors between areas of potential dhole habitat (Rodrigues et al., Reference Rodrigues, Srivathsa and Vasudev2022). As environmental problems increase and financial resources to address them are limited, robust and evidence-based approaches are required to determine priorities for conservation investment (Wilson et al. Reference Wilson, McBride, Bode and Possingham2006). In contrast to previous studies using environmental niche models for the dhole at local to regional scales (Nurvianto et al., Reference Nurvianto, Imron and Herzog2015; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Havmøller et al., Reference Havmøller, Havmøller, Nawangsari, Pratiwi, Møller and Træholt2022; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022), our model encompassed the majority of the species’ range. Although this approach may result in lower predictive accuracy at local scales compared to models trained on more localized data, it enabled us to make large-scale comparisons among regions and countries that could potentially harbour dholes, thus providing important information for guiding future conservation actions for this Endangered carnivore.
We identified most of the potential dhole range in Southeast Asia, which also had a slightly higher average probability of occurrence than mainland China and the Indian subcontinent. However, India contained the largest proportion of potential dhole range amongst the individual countries. India has previously been identified as important for dhole conservation. Kamler et al. (Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015) and Srivathsa et al. (Reference Srivathsa, Sharma, Singh, Punjabi and Oli2020) suggested that the country harbours the largest dhole population. On a smaller spatial scale, large parts of Cambodia, Malaysia and Bhutan are potentially suitable for dholes. These countries, together with Thailand, also have a high relative probability of dhole occurrence. Hence, our study partly agrees with the findings of Tananantayot et al. (Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022), who identified Cambodia, Malaysia and Laos as strongholds of dhole habitat within Southeast Asia, and with Thinley et al. (Reference Thinley, Rajaratnam, Kamler and Wangmo2021), who found that dholes were distributed across all 20 districts of Bhutan. In Indonesia, dholes were historically distributed throughout Sumatra and Java (Kamler et al., Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015), but their distribution on these islands is now much reduced (Havmøller et al., Reference Havmøller, Havmøller, Nawangsari, Pratiwi, Møller and Træholt2022). We found larger areas of potential range in Sumatra compared to Java, where the greater distance to the mainland populations raises further concerns for dhole conservation. Our model identified a limited potential range in mainland China. Dholes have been observed in the north-west of China and occasionally in isolated sites in the Kunlun Mountains, the Karakoram Mountains, the Qilian Mountains and the Altun Mountains during the past 2 decades (e.g. Riordan et al., Reference Riordan, Wang, Shi, Fu, Dabuxilike, Zhu and Wang2015; Xue et al., Reference Xue, Li, Xiao, Zhang, Feng and Jia2015). These observations may represent relict populations that are adapted to arid, semi-arid and alpine habitats from Central Asia to north-west China. Environmental conditions in these habitat types differ greatly from those on the Indian subcontinent and in Southeast Asia, and the demographic responses of dholes to environmental variation, including human persecution, may have differed in these northern regions compared to more tropical areas.
Many forests in Southeast Asia are largely depleted of large mammals because of human persecution (Steinmetz et al., Reference Steinmetz, Srirattanaporn, Mor-Tip and Seuaturien2014; Phumanee et al., Reference Phumanee, Steinmetz, Phoonjampa, Bejraburnin, Grainger and Savini2020). Our model may therefore have identified potential dhole range in forests where the species has been extirpated. For example, a snaring crisis in eastern Indochina (Laos, Cambodia and Viet Nam) has resulted in the recent extirpation of tigers and leopards from these countries despite suitable forests and prey still occurring there (Rasphone et al., Reference Rasphone, Kéry, Kamler and Macdonald2019; Rostro-García et al., Reference Rostro-García, Kamler, Sollmann, Balme, Augustine and Kéry2023). Dhole numbers and distribution in eastern Indochina are also greatly reduced and fragmented because of indiscriminate snaring, and dholes are absent from many parts of this region. Because our model did not consider the impacts of widespread snaring, the potential for dholes to inhabit the potential dhole range identified in eastern Indochina may be limited, at least until the snaring crisis has been resolved. Similarly, because no reliable data are available on prey densities across appropriate spatial scales, we did not include prey abundance in our analyses. We recognize that both human persecution and prey abundance are key variables determining the distribution of carnivores (Dalerum et al., Reference Dalerum, Somers, Kunkel and Cameron2008), including dholes (Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022). However, by not including these variables, environmental niche models can effectively be used to explicitly identify areas where carnivore distribution is limited not by habitat suitability, but by direct persecution or lack of prey (Eriksson & Dalerum, Reference Eriksson and Dalerum2018). Such range limitations require further quantification (Everatt et al., Reference Everatt, Moore and Kerley2019), and we suggest that combining environmental niche models with prey abundance data may yield valuable insights (Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Tananantayot et al., Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022).
The three regions identified as potential dhole range are geographically separated, and our models suggest that habitat in two of the three regions is fragmented. Tananantayot et al. (Reference Tananantayot, Agger, Ash, Aung, Baker-Whatton and Bisi2022) also noted a heavy fragmentation of suitable dhole range within Southeast Asia, and Rodrigues et al. (Reference Rodrigues, Srivathsa and Vasudev2022) made similar observations for India. For species persisting only in small, isolated subpopulations, lack of population connectivity can be detrimental in the long term (Finnegan et al., Reference Finnegan, Galvez-Bravo, Silveira, Tôrres, Jácomo, Alves and Dalerum2021). In South Africa, for instance, it has been recognized that the poor connectivity of subpopulations of the African wild dog Lycaon pictus, which shares many characteristics with the dhole, needs to be addressed to safeguard the species’ future. Consequently, a decision was made to translocate individuals between carefully selected sites to maintain viable subpopulations and create an artificial meta-population (Mills et al., Reference Mills, Ellis, Woodroffe, Maddock, Stander and Rasmussen1998). This conservation intervention has been at least partially successful (Nicholson et al., Reference Nicholson, Marneweck, Lindsey, Marnewick and Davies-Mostert2020), highlighting the importance of maintaining demographic connectivity for species in fragmented landscapes. Although we do not believe that an artificial meta-population approach would be realistic for the dhole across Asia, we suggest that connectivity both between and within regions containing suitable dhole habitat may be critical for the species’ long-term survival. Such connectivity must, by definition, focus largely on matrix habitats outside protected areas, which reiterates earlier suggestions that improving connectivity among population strongholds may yield significant conservation benefits (Prugh et al., Reference Prugh, Hodges, Sinclair and Brashares2008).
Of the evaluated environmental variables, land protection and temperature seasonality were important at both spatial scales. Although the level of complexity in our selected models (i.e. 97 parameters for the coarse-scale and 87 for the fine-scale model) prevents us from drawing any detailed conclusions regarding how these two variables influence dhole distribution, we still regard their importance as informative. Protected land was positively associated with dhole range suitability, and although this relationship may partly have been caused by sampling bias, it does agree with previous suggestions that persisting dhole populations are largely restricted to protected areas (Kamler et al., Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021). As livestock density was also an important variable, human–dhole conflict may be a limiting factor for dhole distribution, similar to the situation for other large carnivores (Srivastha et al., Reference Srivathsa, Sharma, Singh, Punjabi and Oli2020; Thinley et al., Reference Thinley, Rajaratnam, Kamler and Wangmo2021; Ghimirey et al., Reference Ghimirey, Acharya, Yadav, Rai, Baral and Neupane2024). Preserving viable populations of wide-ranging carnivores within protected areas is usually not feasible (Finnegan et al., Reference Finnegan, Galvez-Bravo, Silveira, Tôrres, Jácomo, Alves and Dalerum2021), which further highlights the necessity of focusing dhole conservation on unprotected land. Temperature seasonality also had a strong influence at both scales, but with either non-monotonic or bimodal relationships with dhole range suitability. Temperature seasonality may influence almost all aspects of terrestrial ecosystems (Lisovski et al., Reference Lisovski, Marilyn and John2017), and the observed relationships with range suitability highlight the complex effects climate may have on species distributions. The importance of temperature seasonality suggests that dholes are sensitive to climatic conditions, but the non-monotonic relationship between temperature seasonality and range suitability suggests that local factors such as prey availability and interspecific competition also play a role. The relative importance of the other environmental variables differed between the two spatial scales. The importance of different environmental characteristics as well as the scale dependencies observed in the relative importance of different variables highlight the complexities involved in defining a species’ environmental niche, especially for species with broad niche tolerances.
We recognize that our observation data were biased towards tropical areas, with only a limited number of dhole observations from mainland China. Despite our spatial filtering, our model may thus have under-represented potential range areas in the northern parts of the species’ historical distribution. The bias of observations towards tropical regions could have been caused by field efforts being prioritized in areas where the species is most likely to be observed (Guillera-Arroita et al., Reference Guillera-Arroita, Lahoz-Monfort, Elith, Gordon, Kujala and Lentini2015). The observations we used to train the models may thus reflect at least a large portion of the current dhole distribution, albeit not its full historical range. For instance, Kamler et al. (Reference Kamler, Songsasen, Jenks, Srivathsa, Sheng and Kunkel2015) reported widespread and long-running persecution campaigns against carnivores in the northern regions of dhole’s historical range, and suggested that dholes probably disappeared from large areas of central and southern China during the 1980s and early 1990s. Hence, although our model probably represents a fair quantification of the spatial distribution of areas suitable for the dhole, we propose using regional models for smaller-scale applications. We also suggest that dynamic scale optimization, as used for the brown bear Ursus arctos and snow leopard Panthera uncia (Mateo-Sánchez et al., Reference Mateo-Sánchez, Cushman and Saura2013; Atzeni et al. Reference Atzeni, Cushman, Bai, Wang, Chen, Shi and Riordan2020; but see McGarigal et al., Reference McGarigal, Wan, Zeller, Timm and Cushman2016), may be useful to further improve the spatial accuracy of range predictions for species with broad and plastic habitat tolerances, such as the dhole. We also encourage further studies to quantify the distribution status of dholes in the northern parts of their historical distribution, including China, as well as identifying their ecological requirements in these northern regions.
Apart from the potential sampling bias, some additional caveats apply to our study. Firstly, after appropriate spatial filtering we had a relatively limited sample size, with only c. 1 out of 1,000 cells containing a dhole occurrence. However, MaxEnt has been regarded as robust to limited sample sizes (Wisz et al., Reference Wisz, Hijmans, Li, Peterson, Graham and Guisan2008), and sampling biases associated with spatially unfiltered observations may depress the performance of environmental niche models more than training the models on a more limited number of filtered observations (Boria et al., Reference Boria, Olson, Goodman and Anderson2014). Secondly, our observations included data collected over a period of > 20 years, and there may have been a spatio-temporal mismatch between the observational data and some of the environmental characteristics. However, grouping the observational data into shorter periods would lead to further reductions in sample sizes, which means that models on temporally pooled data are probably the most informative. Additionally, snaring in eastern Indochina has resulted in local extinctions of apex carnivores, including dholes. Therefore, dholes may not occur in seemingly suitable areas because of poaching. Finally, we highlight that the MaxEnt algorithm, just as many other machine learning algorithms, is subject to both conceptual and data-related issues that may cause problems both in model predictions and model interpretations (Araújo & Gusian, Reference Araújo and Guisan2006; Varela et al., Reference Varela, Anderson, García-Valdés and Fernandéz-González2014). We tried to minimize these issues by making biologically justified choices regarding the environmental variables and the model grain. We also used objective criteria in our rigorous model selection approach (Warren and Siefert, Reference Warren and Seifert2011) and in the definition of the cut-off point in the MaxEnt cloglog output that delineated potential range. We therefore believe that our modelling process was based on biologically relevant information and objective analytical criteria, as far as this was possible with the information available.
To conclude, we identified potential dhole range in three disparate regions, and connectivity appeared limited both between and within these regions. Hence, we suggest that conservation actions should be focused on activities within each of these three regions, and on improving connectivity amongst dhole populations. As the majority of the potential dhole range was identified in Southeast Asia, and countries within this region also had a higher proportion of their total land area identified as potential dhole range, this region should be a priority for dhole conservation. However, amongst individual countries, India harbours the highest proportion of potential dhole range, which agrees with previous suggestions that the country probably also harbours the largest proportion of the global dhole population. Coordinating conservation efforts between regions in India and Southeast Asia could thus be a key aspect of future dhole conservation planning. We encourage transboundary conservation initiatives integrating areas in southern China, Myanmar, north-east India, Nepal and Bhutan. Our study also highlights the need for more monitoring and assessments of dhole population status and restoration potential in the northern parts of its historic distribution, including in mainland China. Finally, we suggest that focusing dhole conservation on population persistence in unprotected areas may be key to ensure the long-term viability of this species, both by improving connectivity amongst highly suitable habitat patches but also by avoiding problems associated with efforts to maintain viable populations of wide-ranging species within restricted protected areas.
Author contributions
Study conceptualization: MPK, FD, KK, WW; data collection: all authors; data analysis: MPK, FD; writing: MPK, FD, KK, WW; revision: all authors.
Acknowledgements
We thank the numerous people and organizations that collected the dhole presence data. These data were compiled during a workshop organized by the dhole working group of the IUCN Species Survival Commission Canid Specialist Group, the IUCN Species Survival Commission Conservation Planning Specialist Group, Kasetsart University in Thailand, and the Smithsonian Conservation Biology Institute. The workshop was funded by Columbus Zoo and Aquarium, Copenhagen Zoo, Minnesota Zoo, San Diego Zoo Global and the Smithsonian Conservation Biology Institute. The Wildlife Conservation Network provided a Sidney Byers scholarship to MPK, the Spanish National Research Council provided funding to MPK, APK and FD (COOPB23009), the Spanish Ministry of Economy and Competitiveness provided a Ramon y Cajal fellowship to FD (RYC-2013-14662), and KMPMBF was funded by a research grant from the Conselho Nacional de Pesquisa e Desenvolvimento Científico e Tecnológico (CNPq, 308632/2018-4). The collection of dhole data from Nepal was supported by grants from The Rufford Foundation (44630-D, 14005-B, 11636-2, 8939-1, awarded to APK), an award to APK by the People’s Trust for Endangered Species and a grant by Conservation Connect, an initiative of Prince Bernhard Nature Fund, awarded to AACD.
Conflicts of interest
None.
Ethical standards
This research abided by the Oryx guidelines on ethical standards.
Data availability
The predicted rasters from the MaxEnt models are available in cloglog format together with the thinned observations used to train the coarse- and fine-scale models, as well as scaled and aligned environmental layers used for each model, on the Figshare platform (doi.org/10.6084/m9.figshare.29141738).
Open access
