Ecological niche modeling of rare, endangered, and endemic species poses serious challenge to researchers (Engler et al. 2004; Pearson et al. 2007; Lomba et al. 2010) due to absence of appropriate occurrence data. Recent studies, indicate that species with restricted range distribution or narrow biological niches perform better as their specific requirement is easy to define in a multivariate setting (Segurado et al. 2004; Luoto et al. 2005; Coetzee et al. 2009; Marimon et al. 2009). The current research of building a robust niche model with 21 spatially independent points distributed across Western Ghats and adjoining areas. The attempt aimed at few key parameters for information deficient endemic taxa to develop baseline distribution map for future inventory and conservation plan (Engler et al. 2004; Siqueira et al 2009; Marini et al 2010).

In data deficient studies, ensemble method provides opportunity to assess performance of multiple algorithms in a common limiting condition. In this study, statistical performances of the ensemble sets are quite satisfactory, but projected distributions have shown much variation than expected. The binary maps created from ROC value of ensemble weighted mean models varied considerably  across algorithms. Both over prediction and over-fitting are observed in CTA-RF, GLM-GAM-ANN techniques due to the limitations of the algorithms and predictor variables to deal with the limited occurrence points. However, validation in consultation with the regional domain experts indicate that ensemble output from MARS has shown near realistic distribution pattern. In addition, MaxEnt generated relatively better realistic map than ensemble method confirming to earlier findings on its appropriateness (Pearson et al. 2007; Graham et al. 2008; Wisz et al. 2008; Mateo et al. 2010). 

It is due to inadequate distribution data, defining confirmed absence points have been a problem; paving the way for pseudo-absence (pa) strategy. Multiple options such as  number of pa points, sets of pa to be included, and pa selection strategies were examined. Algorithm specific pseudo-absence point sets performed better than uniform pseudo-absence points (i.e. 10,000 points for all algorithms) (Fig 4). On the contrary, model performance has not differed much when pseudo-absence selection strategy was evaluated i.e. geographical exclusion (5km buffer) vs. random selection. Final run with three pa sets has been selected after initial trials with different sets have not shown much change in performance (in terms of TSS, Kappa and ROC values beyond three sets).

4.1 Ecological implications of the models
Out of seven selected variables, temperature (bioclim 7, 10 and 16) and wetness (bioclim??) have profound influence on species distribution. The result corroborates the general understanding of the species’ preferential distribution in perennial or seasonally wet regions (Sasidharan 1997; Chandran et al. 2008; Roby et al. 2013). Its clustered presence near southern Western Ghats is perhaps related to high wetness with the longer period (8-10 months) of rainy days and higher quantum of rainfall (Prasad et al. 2008; Anu et al. 2009). Higher number of rainy days influences seasonality and isothermality while keeping temperature differences minimal. The scattered distribution of suitable areas in central and northern Western Ghats may be due to a decline in species habitat (with suitable microclimate), variations in rainfall (there is a gradual northward decline of rainfall from south both in terms of duration and quantity), and large scale forest fragmentation in addition to inadequate exploration.   

Figure 5. Potential distribution map of S.travancoricum Gamble. (a = MARS ensemble output, b = CTA and RF ensemble output, c = GLM,GAM and ANN ensemble output and d= MaxEnt output )

4.2 Conservation status assessment
The extent of occurrence was examined through two widely used methods which have demonstrated almost similar estimation for the species. On the contrary, potential distributions obtained from modelling experiments have shown wider variation. The area obtained through MaxEnt was  relatively closer to the EOO values estimated through IUCN recommended protocols than the other values (Table 2). The difference between the outputs could be attributed to the applied methods (Sergio et al. 2007); while conventional EOO estimation is based on topological methods, niche models rely on spatial as well as environmental space, but disjunctions are equally considered in both the cases. Similarly, in AOO estimation the scale of the grid is an important factor; however, following IUCN recommended procedure the estimated area of occurrence (i.e. 116 km2) has far exceeded the threshold for critically endangered species of 10 km2.

Application of niche modeling in EOO and AOO estimation had earlier generated diverse responses. Sergio et al. (2007) and Alfaro et al. (2012) have identified ENM based status assessment as ecologically robust and realistic; in contrast, Attorre et al. (2013) and Syfert et al. (2014) have emphasized that the spatial distribution of the species and available information on their distribution play a major role in EOO estimation. In the current investigation, modelling outputs were also varied based on the algorithms, especially for commission and over fitting errors. Moreover, S. travancoricum has been reported from Southern Western Ghats swampy areas for a long time (Bourdillon 1908; Gamble 1935), whereas report of central Western Ghats and northward distribution is relatively recent (Chandran et al. 2008, 2010; Prabhugaonkar et al. 2014; Ray et al. 2012, 2014). A close observation on the occurrence points can detect the gap between southern and central Western Ghats, which can be ascribed to the dearth of exploration or to the changes in the rainfall pattern. Both modeling and EOO studies have incorporated this gap area as a possible place for future inventory for finding new populations.

Based on these findings, a reassessment of species’ position in IUCN list is essential, as the current EOO estimates are much larger than the < 100 km2 threshold for critically endangered species, putting it under least concern category. Likewise, the magnitude of AOO is also greater than the threshold (< 10km2) according to IUCN guideline, suggesting endangered status of the species. The current results are in conformity with the earlier study (Roby et al.,2013), emphasized the need to consider the species under a changed criterion (from C2a to A1e).  Importantly, the current IUCN status of the plant has been assigned based on CAMP workshop study in 1998; thus requiring a thorough recent revision (IUCN 2015). Future exploration in suitable areas to discover new populations may be an integral part for review of the distribution range and population status, threats to the habitats, and other associated parameters (Santos et al. 2006; Papes and Gaubert 2007; Brito et al. 2009).