The clouded leopard Neofelis nebulosa, categorized as Vulnerable on the IUCN Red List, is threatened throughout its range, primarily because of habitat loss and fragmentation, prey exploitation and poaching (Gray et al., Reference Gray, Borah, Coudrat, Ghimirey, Giordano and Greenspan2021). While viable populations may remain in some regions, others have undergone catastrophic declines and the species has probably been extirpated from Viet Nam, most of China and large parts of Cambodia and Laos (Petersen et al., Reference Petersen, Steinmetz, Sribuarod and Ngoprasert2020). Another region that has experienced extirpation of the clouded leopard is the island of Taiwan. During 1997–2012, an extensive camera-trap survey of Taiwan's mountainous interior, believed to be the species' last stronghold on the island, did not find any evidence of the felid, and it was concluded the species had been extirpated decades earlier, probably as a result of deforestation and overexploitation of both the clouded leopard and its prey (Chiang et al., Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015). However, populations of many of Taiwan's mammals are now recovering (Weng et al., Reference Weng, Liu and Tuanmu2023), mainly because of an increase in forest cover across the island (Chiu et al., Reference Chiu, Huang, Wu and Hsieh2015) and the prohibition of hunting in the 1970s (Sun et al., Reference Sun, Arora, Lin, Lin, Chi and Chen2019). Encouraged by this and a positive attitude regarding the return of the species among Taiwanese citizens (Greenspan et al., Reference Greenspan, Giordano, Nielsen, Sun and Pei2020), a team led by the Clouded Leopard Association of Taiwan and supported by international experts from Panthera, the IUCN Species Survival Commission Cat Specialist Group and The Wildlife Conservation Research Unit is considering a clouded leopard reintroduction programme in Taiwan. However, the programme is hampered by a lack of analysis of potential suitable clouded leopard habitat on the island.

To date, only one attempt has been made to quantify available habitat for the clouded leopard in Taiwan. Chiang et al. (Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015) identified areas of suitable habitat for the clouded leopard based on prior knowledge of the species’ habitat requirements. However, without empirical data on the species itself, the habitat assessment could be biased. Thus, to examine the availability of suitable habitat for reintroduction, we applied a multi-scale multivariate habitat selection model, developed using extensive camera-trap records from across the species' range (Macdonald et al., Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019). This allowed us to predict and quantify potential habitat for the clouded leopard in Taiwan.

We focused on the main island of Taiwan, which historically encompassed the most easterly part of the clouded leopard's range. At c. 36,000 km², Taiwan is a relatively large, rugged and mountainous island that supports a diverse range of habitats, from lowland tropical forest to alpine grassland. Despite significant development in the lowland and coastal regions, over 60% of the island is classified as forest (Chiu et al., Reference Chiu, Huang, Wu and Hsieh2015) and large swathes of continuous forest persist in the upland interior.

To predict suitable habitat for the clouded leopard in Taiwan, we extrapolated a habitat suitability model developed for the species' extant range (Macdonald et al., Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019), hereinafter referred to as the empirical model. This empirical model was developed using binary presence–absence data of clouded leopards derived from 2,948 camera-trap stations across 45 study sites in nine countries, spanning the full range of N. nebulosa throughout South and Southeast Asia. Macdonald et al. (Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019) analysed these data with a multi-scale optimization approach (sensu McGarigal et al., Reference McGarigal, Wan, Zeller, Timm and Cushman2016) with a generalized linear mixed model to predict clouded leopard habitat suitability. The empirical model, which averages four models with ΔAICc ≤ 2 (difference in Akaike information criterion corrected for small sample size, compared to the best performing model), incorporates nine environmental variables: (1) per cent of closed forest (forest cover > 40%), (2) mean compound topographic index (i.e. flow accumulation, with lower values in higher elevation areas), (3) mean annual precipitation, (4) per cent of mosaic land cover (i.e. mixed land cover types), (5) correlation length (i.e. the average distance an individual can travel within habitat patches) of protected area, (6) correlation length of grassland/shrubland, (7) standard deviation of slope, (8) mean of slope, and (9) per cent of tree cover. All covariates are defined in Macdonald et al. (Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019).

We hypothesized that clouded leopard habitat quality in Taiwan would be influenced by similar factors to those that drive the species’ distribution elsewhere. We applied the empirical model to Taiwan's environmental conditions, using a combination of the same environmental factors (Table 1) to map predicted suitable habitats based on the species' known habitat preferences elsewhere. As there are few records of clouded leopard occurrence above 3,000 m, we removed areas that exceeded this elevation, as they are likely to be used infrequently. To assess the appropriateness of using the empirical model to extrapolate beyond its original geographical extent, we visually inspected the multivariate environmental similarity surface (MESS; Elith et al., Reference Elith, Kearney and Phillips2010) developed by Macdonald et al. (Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019), which provides an index of similarity between the conditions at sampling locations used for model development and the environmental space in Taiwan.

Table 1 Camera-trap days and combination of nine environmental variables included in the empirical model (Macdonald et al., Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019) when used to project the suitable habitat for the clouded leopard Neofelis nebulosa in Taiwan, with the optimal scale, Akaike information criterion (AIC) importance, averaged and adjusted standard error β (coefficient) for four models, z, and P-value of each covariate. (The AIC importance is the importance of a covariate in improving the model when it is included. It is calculated as the sum of the Akaike weights for all models that include the covariate.)

To provide further insight into the potential for clouded leopard reintroduction in Taiwan, we grouped the continuous habitat suitability surface into three classes, using a range of increasingly stringent thresholds of suitability: unsuitable (< 75th percentile), moderately suitable (≥ 75th percentile) and highly suitable (≥ 90th percentile). Lastly, we compared the suitable habitat with (1) Chaing et al.'s (Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015) prediction of clouded leopard habitat, which relied on expert opinions, hereinafter referred to as the expert model, and (2) the protected areas network in Taiwan (UNEP-WCMC, 2024).

Our predictive model suggests that a large, continuous area of 13,854 km² (38% of the island's landmass) in Taiwan's hilly and mountainous interior, is composed of moderately (10,041 km²) and highly (3,813 km²) suitable habitat for the clouded leopard (Fig. 1a; Table 2). Inspection of Macdonald et al.'s (Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019) multivariate environmental similarity surface indicates that the multivariate environmental space used for training of the empirical model was similar to the environmental conditions in areas of predicted suitable habitat in Taiwan, suggesting that extrapolation of this model to Taiwan is viable. Our findings corroborate Chiang et al.'s (Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015) expert model, with areas they predicted to be suitable habitat also being predicted as suitable by our model. However, our model predicted a much larger area of moderately to highly suitable habitat than the expert model, which estimated suitable clouded leopard habitat to be only 8,523 km² (Fig. 1b). Intersection analysis shows that 46% of the area predicted by our model as suitable for the clouded leopard is under protection, of which 3,419 km² (90%) and 2,940 km² (29%) are highly and moderately suitable habitats, respectively (Fig. 1c).

Fig. 1 (a) Predicted suitable habitat for the clouded leopard Neofelis nebulosa in Taiwan. (b) Comparison of suitable habitat predicted by applying the empirical model (Macdonald et al., Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019) to Taiwan and by the expert model (Chiang et al., Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015). (c) Protected areas network in Taiwan (UNEP-WCMC, 2024) and suitable habitat predicted by application of the empirical model.

Table 2 Highly and moderately suitable habitat predicted by the empirical model when applied to Taiwan, with the predicted suitable habitat within protected areas in Taiwan, the predicted suitable habitat that overlaps with the expert model (Chaing et al., Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015), and the total predicted suitable habitat as a per cent of the total land area of Taiwan.

Our extrapolation of an optimized empirically based habitat model for the clouded leopard on the Asian mainland to the island of Taiwan predicted extensive and connected habitat within Taiwan's rugged, mountainous terrain. Additionally, our habitat suitability predictions largely align with the expert model for the areas of highest suitability but suggest that the suitable habitat could be more extensive than predicted by the latter. The difference in predicted suitable habitat between models could result from different perspectives regarding anthropogenic impacts on the species. Chiang et al. (Reference Chiang, Pei, Vaughan, Li, Chen and Liu2015) classified several areas as unsuitable based on the proximity to roads and human settlements. However, considering the uncertain influence of road proximity on the species (Kaszta et al., Reference Kaszta, Cushman, Htun, Naing, Burnham and Macdonald2020), we included suitable habitats even if they were close to roads. Although hunting pressure is known to affect the species' distribution on the mainland (Petersen et al., Reference Petersen, Steinmetz, Sribuarod and Ngoprasert2020), hunting pressure in Taiwan is likely to be low because of the country's long-standing and effective hunting bans (Sun et al., Reference Sun, Arora, Lin, Lin, Chi and Chen2019).

Our results show that 90% of the predicted highly suitable clouded leopard habitat in Taiwan is protected, exceeding the protection rates of most core habitats in Southeast Asia (Macdonald et al., Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019). When also including moderately suitable habitat, 46% of all suitable habitat is under protection, still above the Southeast Asian mainland average of 32%. Taiwan's substantial habitat protection, coupled with low hunting pressure, suggests it could be suitable for reintroduction of the clouded leopard.

It is possible, however, that extrapolating Macdonald et al.'s (Reference Macdonald, Bothwell, Kaszta, Ash, Bolongon and Burnham2019) empirical model to Taiwan could overestimate the extent of suitable habitat because it subsumes continental scale variation into a single global model, ignoring potential nonstationarity in local limiting factors and realized habitat niche (i.e. island and continental difference in climate dynamics or sympatric species; e.g. Cushman et al., Reference Cushman, Kilshaw, Campbell, Kaszta, Gaywood and Macdonald2024a,Reference Cushman, Kilshaw, Kaszta, Campbell, Gaywood and Macdonaldb,Reference Cushman, Kilshaw, Kaszta, Campbell, Gaywood and Macdonaldc,Reference Cushman, Kilshaw, Kaszta, Campbell, Gaywood and Macdonaldd). Nonetheless, statistical extrapolation of a habitat model developed from a large empirical database is probably the best method currently available to assess habitat potential for the clouded leopard in Taiwan. Finally, this study is preliminary and limited with respect to prey distribution and abundance, functional connectivity and social feasibility, all of which should be considered in the development of more refined plans for this proposed reintroduction. Our study is the first empirically-based attempt to predict habitat suitability for the clouded leopard in Taiwan and will be of value to develop strategies and garner support for the reintroduction of this iconic species.