Data preparation

We collated unpublished Javan leopard points of occurrences recorded by various field research projects and individuals between 2008 and 2014 (S1 Table). For security reasons, we present the approximate location of each Javan leopard locality within a 100 km² rectangle (S1 Fig). We then selected points that had verifiable evidence, including camera trap videos and photographs, human-leopard conflict reports, pugmarks, scratches and feces. Given the absence of other large cats in Java, we believed that misidentification of these signs to be unlikely.

We used ten environmental variables that have been shown to be suitable predictors of large carnivore presence by other studies. These included distance from roads [23], distance from forest edges to the interior and exterior, and protected areas [26], elevation [23,26], distance from rivers [27], slope [23,28], land use classification, precipitation, and temperature [29]. The original spatial dataset included protected area land use classification layers [30], roads, rivers and Indonesia baseline maps [31], digital elevation map (30 m resolution) [32], and mean precipitation and temperature (1 km resolution) [33]. We converted the protected area layer into a binary raster containing '0' (non-protected area) and '1' (protected area).

The original land use classification layer contained 18 classes. For better interpretation, we reduced it to 15 classes based on similarities and converted them into raster layers. We generated the slope layer from the digital elevation map. We selected forest polygons from the land use classification and generated Euclidian distances to produce the distance from forest edge to the Javan leopard points recorded both inside and outside of the primary forest polygons. Thus, for distance from forest edge to the exterior, all Javan leopard points that fell within the forest were assigned a value of '0' and vice versa for the distance from forest edge to the interior. Similarly, we employed Euclidian distance to roads (vector) and rivers (vector) to generate the distance from roads and rivers (30 m resolution) layers, respectively.

The primary analysis tool used in this study required all datasets to have exactly overlapping cells and spatial extent [34], which in this study was confined to the island of Java. We generated a raster mask of 0.25 km² cell size covering Java to provide a baseline environment setting for further resample processing of the background layers. We then re-sampled the Javan leopard points and all environmental variables with the application of a mask layer. This procedure removed any duplication of Javan leopard points within a cell for subsequent analysis. Because several environmental variables were re-sampled from finer to coarser resolution, we performed a bilinear interpolation technique to assign a new value to a cell by using a weighted distance average of four adjacent input cells. All spatial processes were performed using Raster Processing and Spatial Analyst tools in the software ArcGIS 10.2 (ESRI, Redlands).

The Javan leopard points were found to be biased toward the sampled areas, a situation common to presence-only datasets [35]. To control for this sampling bias, we converted the mask layer into a bias grid [36]. In the analysis, the value of a cell (c1) in the bias grid was used to assign greater weight to Javan leopard points (c2) with fewer spatial neighbors and vice versa. The value of c1 was a sum of the distances between c1 and c2 as calculated using the Gaussian Kernel function, w = exp(−d²/2s²), where w is the weight, d is the distance (in km) between c1 and c2, and s is the standard deviation. We used 4.2 km for s because it represents the known diameter of the largest Javan leopard home range size (13.6 km²) [18,19]. Following the Gaussian distribution, s, yielded points that are, for example, located 4.2 and 8.4 km away from a cell as having 60.6% and 13.5% as strong an influence, respectively. We used the Distance Among Points tool of the Geospatial Modeling Environment version 0.7.3.0 [37] and the Calculate Variable and Aggregate tools of the IBM SPSS Statistics software version 20 (IBM Corporation 2011) to calculate the distance between c1 and c2 and the Gaussian Kernel function, respectively.

Data analysis

We predicted suitable landscape of Javan leopard using Maxent version 3.4.1. Maxent has been widely used for presence only dataset over other techniques due to its robustness against autocorrelated environmental predictors [38,39], lower sensitivity to small sample sizes [40], and being less affected by spatial errors [41]. The final sample inputs consisted of 169 Javan leopard presence points. We performed a Pearson’s correlation analysis using the Hmsc package within R software (R Development Core Team 2010) to test for correlations between the ten environmental variables, from which a pair of variables was removed if the coefficient correlation was > 0.50 [42]. The final set of uncorrelated environmental variables used in the subsequent analysis included distance from forest edges to the interior and exterior, distance from roads, protected areas, elevation, distance from rivers, land classification and mean precipitation.

We set the protected areas and land classifications as categorical variables and the rest as continuous variables. We performed a Bootstrap procedure with 25% random tests, ten replicates, and 5,000 iterations, and kept the other settings at the default option. We also performed Jackknife tests to assess consistency in variable importance between the training and test gains [43]. The overall model performance was measured by the area under the curve (AUC) of the receiver operating characteristic (ROC) curve [39]. We estimated the relative importance of each predictor to the Maxent model using the percent contribution and permutation importance, averaged over ten replicates. We investigated the response curves to explore how the environmental predictors effected the Maxent prediction.

Identifying landscapes and defining its characteristics

We used the model prediction to determine suitable Javan leopard landscapes based on a ten percentile training presence logistic threshold. Model pixels with logistic probabilities smaller than the threshold were omitted from the final predictive model [24,34,44]. We converted the predicted suitable patch raster into a polygon format and defined the minimum suitable patch size for Javan leopard as being large enough to contain at least five mature individuals [28]. Subsequently, we retained all predicted suitable patches of at least 68 km² or equal to five times the largest known Javan leopard home range [18,19]. For a Critically Endangered subspecies like Javan leopard, smaller suitable patches with confirmed evidence of Javan leopard may still be important if it supports connectivity between patches. We therefore retained all smaller suitable patches with and/or close to Javan leopard localities up to 4.2 km away or equal to the diameter of their home range. Predicted suitable patches that did not meet these criteria were removed from further analyses.

We assumed all protected areas that fully or partially overlapped with the predicted suitable patches as being important for Javan leopard. We updated these patches with protected areas using the Update tool of the ArcGIS 10.2. This step removed all stand-alone protected areas and combined them with associated suitable patches into one unique polygon. We calculated the area of the identified suitable patches in each protected area as a surrogate indicator of the ecological importance of each protected area to support Javan leopard.

We included the protected area layer in the land use classification layer, so that it would be considered in the analysis. This step removed all land classes that fell within a protected area into one unique polygon coded as “protected area” and produced 16 land use classes. We defined suitable landscapes as patches that: 1) are predicted to be suitable for Javan leopard based on a ten percentile training presence logistic threshold, 2) are larger than 68 km², and/or 3) have evidence of Javan leopard presence, and/or, 4) within 4.2 km of a Javan leopard locality. We then extracted the land use classes in each suitable landscape by intersecting the landscape layer with a land use classification layer using the ArcGIS 10.2 Intersect tool.

Study limitations

A limitation of this study was that it was unable to incorporate a data layer on prey availability due to the absence of such data from the island of Java [22]. It was not, therefore, possible to investigate the potential of the predicted suitable landscapes, especially those are outside the protected areas and forest habitats, to support viable populations of Javan leopard over the long-term. Future studies should therefore aim to address this question, especially in the priority landscapes. The Javan leopard occurrence data used in our analysis might have certain limitations that are associated with the use of a presence-only dataset. First, data were collated from multiple sources and not based on surveys that used the same sampling method. However, because such a dataset does not exist, we therefore aimed to collect as much data as possible and then filter these to remove redundancy whilst maintaining a comprehensive dataset. Second, although Maxent is less sensitive to small sample sizes [40] and less affected by spatial errors [41], it requires that the samples (occurrence data points) be unbiased and therefore independent of the distribution of the target species [43]. The high AUC value recorded in our model might be an artifact of the AUC statistic, which tends to be higher for species with small home range sizes relative to the study area [43]. However, the potential sampling bias here might have been overcome by our study using a relatively large dataset recorded from all possible habitat types across Java. Furthermore, we controlled for potential sampling bias by using bias grids as a function of their Gaussian Kernel distribution and overcame the dependency issue by resampling the original Javan leopard records into a density of one point per 0.25 km².

Characterizing Javan leopard landscapes

Our study provides fresh thinking into the role of modified landscapes surrounding protected areas for improving the carrying capacity and, ultimately, long-term survival of the Javan leopard. The Maxent model shows higher association between leopards and more productive landscapes in western Java than the drier forests in eastern Java. This was revealed by the sharp increase in the probability of Javan leopard occurrence in landscapes with higher precipitation. Thus, higher precipitation should support higher plant productivity for the ungulate prey base [45]. However, with the absence of prey data, we were not able to further evaluate the relationship between precipitation and prey base availability. The low contribution of the protected area variable to the overall model output might be explained by the fact that most protected area boundaries were located within larger forest areas, therefore outperformed by the distance from the forest habitats variable.

We identified 70% of the protected area network in Java as providing suitable leopard habitat. However, this network contributed only 40% of the total area of the suitable landscapes. A population viability analysis on ten leopard populations in South Africa predicted that no population with fewer than 50 individuals would survive over 100 years without dispersal and connectivity to a neighboring population [46]. Thus, based on a maximum estimated Javan leopard density, 16 leopards/100 km² [19], 12 suitable landscape patches (#1, 4, 6, 7, 9, 14, 16, 18, 23, 24, 25, 26) were considered large enough to support at least 50 individuals. Here, Javan leopard recovery efforts should focus on strengthening protected area management to reduce poaching and maintain landscape integrity, before moving on to options such as connecting landscapes and their respective leopard populations.

From the landscapes identified with good recovery potential, it is important to stress that, besides the national parks, all other protected area types were found to be too small on their own to hold viable populations of Javan leopard. This situation was also found in several studies in West Africa and North America where many wildlife reserves were to be too small or had inadequate natural resources for achieving large carnivore conservation objectives [47,48]. For the Javan leopard, this highlights the importance of managing suitable landscapes that adjoin these protected areas so as to support viable populations. However, these buffer zone habitats may represent poorer quality habitat and have a lower prey base and therefore support a lower density and number of leopards. Determining leopard densities in these areas and, therefore, how large these areas should be is a topic for future research.

National Park is listed as a Category II protected area of the World Conservation Union [49] and is the strongest protected area type afforded by the Government of Indonesia, in terms of infrastructure, management, financial support and human resources allocated. Thus, strengthening national park management should be considered as a top priority for the long-term survival of the Javan leopard. We identified only three national parks with sufficient habitat to potentially support more than 50 individuals: Gunung Halimun Salak (4, 74,018 ha); Ujung Kulon (1, 42,910 ha); and, Meru Betiri (#26, 42,465 ha). Among these, only Gunung Halimun Salak could potentially support more than 100 individuals, which underlines its importance as a flagship protected area. Two other national parks, Gunung Gede Pangrango (#6) and Bromo Tengger Semeru (#24) could potentially support 50 leopards or more if suitable habitat in the adjacent areas is included.

Understanding the wider landscape characteristics is vital to increasing the population size of large carnivores that are threatened with extinction [50,51]. This not only includes the identification of suitable landscape adjacent to the protected areas, but the key stakeholder groups who should be engaged by conservation managers. Our study found that more than half of the suitable landscape outside of the protected areas is in production forest and secondary forest. Further, with less than 5% of primary forest occurring outside of protected areas, production and secondary forests should therefore play an important role in providing additional habitat for the remaining Javan leopard populations. Here, Government Regulation No. 72/2010 mandates the management of non-conservation state forest in Java to Perum Perhutani, a company under the Ministry of State-owned Enterprises, as the main investor, and the Ministry of Environment and Forestry, as an advisor for its technical and operational activities.

Beside state forest, the Government of Indonesia recognizes the legal right of local communities to manage private forest, which is commonly known as hutan rakyat (Community Forest). The role of local communities in managing hutan rakyat adjacent to state forest is legalized by Perum Perhutani through the establishment of a Community Village Forest Institution (Lembaga Masyarakat Desa Hutan). Under this scheme, local communities are permitted to develop agroforestry for commodities, such as teak, acacia, silk trees and mahogany. In 2009, 2.7 million ha of hutan rakyat in Java was managed by 690,895 households through agroforestry schemes [52,53]. Most hutan rakyat is situated adjacent to state forests and may therefore contribute significant Javan leopard habitat that is located around protected areas.

This study confirms the resilience of leopard outside of its main forest habitat type [54]. We found that nearly half of the Javan leopard data points were recorded outside of protected areas and primary forest, which concurs with other studies that found a high level of leopard adaptability in modified habitats [4,7,11,12,24]. This indicates the ability of leopards to subsist in and move through modified habitat [13], as has been found in India [55], South Africa [56] and Russia [57]. These modified habitat may therefore serve as structural corridors that facilitate dispersal to enable source-sink connectivity between viable and otherwise non-viable leopard populations [46]. Balme et al [58] recorded leopards moving far beyond the productive natural habitats and into areas where they were then killed, either deliberately or accidentally, by people. A successful strategy for conserving a wide-ranging large carnivore will therefore rely on protecting the source population and providing dispersal opportunities with sink populations through maintaining connectivity [24,29,59,60]. This should be conducted with a reduction in leopard and prey offtake from hunting, retaliatory killing and problem animal removal [46]. Further studies on Javan leopard movement and home range size as a function of prey availability and potential threats in different habitat types, especially human-modified ones, will allow conservation managers to better understand the species’ response to available resources and different threats, and thus their survival and reproduction potential [61].

More than a quarter of our Javan leopard occurrences were outside of the protected area network and from conflicts with local communities. For a Critically Endangered subspecies, this is of great concern because of the risk posed by retaliatory killings to conflict incidents can have a disproportionately large impact on small population sizes. These types of attacks are well documented in causing negative local community perceptions and attitudes towards the conservation of large carnivores [62–64]. For Java, the history of the tiger’s extirpation from the island approximately 30 years’ ago provides a salutary lesson for the future of the Javan leopard because ultimately the loss of habitat and ensuing competition with people for space and resources, particularly its ungulate prey base, was to the tiger’s great detriment [65]. The Javan leopard, despite being more adaptable than the tiger, now occupies an island that whilst only accounting for 6.9% of Indonesia’s land mass is home to 60% of its human population. This restricts core forest habitat to volcanic peaks and coastal corners. Despite this, our study identifies that many of the occupied landscapes may still contain viable leopard populations, making their protection a top priority. It also identifies smaller landscapes that should be connected to achieve the same aim. Thus, our spatially-explicit modelling provides time-critical information for implementing the 2016–2026 Javan leopard action plan to better effect. This approach could also be adopted for other Critically Endangered species for which there is sparse data, such as the saola.