Local extinction of a montane, endemic grassland bird driven by landscape change across its global distribution

Context Tropical montane habitats support high biodiversity, and are hotspots of endemism, with grasslands being integral components of many such landscapes. The montane grasslands of the Western Ghats have seen extensive land-use change over anthropogenic timescales. The factors influencing the ability of grassland-dependent species to persist in habitats experiencing loss and fragmentation, particularly in montane grasslands, are poorly known. Objectives We studied the relationship between the Nilgiri pipit Anthus nilghiriensis, a threatened endemic bird that typifies these montane grasslands, and its habitat, across most of its global distribution. We examined what habitat features make remnants viable habitat, which is necessary for effective management. Methods We conducted 663 surveys in 170 sites and used both single-season occupancy modelling and N-mixture modelling to account for processes influencing detection, presence, and abundance. Results Elevation had a positive influence on species presence, patch size had a moderate positive influence and patch isolation a moderate negative influence. Species abundance was positively influenced by elevation and characteristics related to habitat structure, and negatively influenced by the presence of invasive woody vegetation. Conclusions The strong effect of elevation on the highly range-restricted Nilgiri pipit makes it vulnerable to, and an indicator of, climate change. This highly range-restricted species is locally extinct at several locations and persists at low densities in remnants of recent fragmentation, suggesting an extinction debt. Our findings indicate a need to control and reverse the spread of exotic woody invasives to preserve the grasslands themselves and the specialist species dependent upon them.

suggested that the Nilgiri pipit was restricted to areas above 1900 metres 144 above sea level (a.s.l.). Conservatively, we limited our survey to grasslands above 1600m a.s.l., 145 and also confined our study to areas in which verifiable contemporary records of the Nilgiri pipit 146 exist. This region encompasses the two major high-altitude plateaux of the Western Ghats; the 147 Nilgiris and the Anamalai-Palani Hills. Although the species has been reported elsewhere, 148 photographic evidence or capture records for these locations, do not exist and intensive surveys across the smaller northern and southern grasslands have failed to detect the species (Robin & 150 Sukumar, 2002;Robin et al., 2006;Sasikumar, Vishnudas, Raju, Vinayan, & Shebin, 2011). Within the above area, we mapped the extent of montane grasslands using Sentinel-2A imagery. 158 We obtained imagery from the USGS Global Visualization Viewer (GloVis; 159 https://glovis.usgs.gov/). Satellite imagery were obtained from the dry season (February 2017), 160 when cloud cover was low. We removed atmospheric components such as dust particles, water 161 vapour, and atmospheric temperatures in the satellite images by generating ground reflectance 162 images using the Sen2Cor processor in SNAP v. 5.0.8 (ESA, 2017). We used a combination of supervised and unsupervised classification to map the montane grasslands (see supplementary 164 methods for further detail). The overall accuracy of this classification, determined using 100 165 ground-truth GPS locations across the entire study area, was 96.5%, while the Kappa coefficient 166 was 0.93 (following Congalton, 1991).

168
The final selected area represented 434.98km 2 , or 85%, of the 511km 2 of grassland above 1600m 169 in the Western Ghats. It consisted of 1,449 discrete grassland patches, varying in size from <1ha 170 to 120,000ha. Grassland patches between 4 and 25ha were designated as sampling units, 171 encompassing the range of our estimates for Nilgiri pipit home range (Vinod 2007, Vinod, U., 172 personal communication, 2018. We placed patches between 1 and 4ha into clusters, if each 173 patch was within a maximum of 200m (based on available information about pipit movement; 174 Vinod, U., personal communication, 2018) away from at least one other, ensuring that slightly 175 fragmented grasslands which were effectively larger than 4ha were not discarded. We 176 determined the total area of each cluster, and discarded clusters or individual patches totalling 177 less than 4ha as these were unlikely to support the species' occurrence. We laid a 500m grid 178 across all grassland patches larger than 25ha, designated each grid cell a separate patch, and 179 removed all patches smaller than 4ha. We randomly selected 202 sites from the remaining 2,378

183
To accurately model both detection and occupancy, we visited each site multiple times (Max. 184 visits = 4: 11 sites visited thrice, one site twice, one site once), with the duration of each visit set 185 proportional to site area: species detection probability was thus equal across patches of variable size. We expended 2 minutes of survey effort per hectare, which we determined to be optimal 187 based on reconnaissance surveys. Thus, surveys were between 8 and 50 minutes long. Surveys 188 were conducted on foot; Nilgiri pipit counts were recorded based on visual and auditory 189 detections strictly within the sampling site. Covariates were recorded at the site-level and the 190 visit-level (Table 1:    Species-level detectability effects include individual-level detectability effects and abundance effects. ++ indicates a strongly positive expected effect; + indicates a positive expected effect;indicates a negative effect: N indicates no expected effect. *For grass height, occupancy and abundance were expected to be highest for the intermediate category. ‡Detectability was expected to be low in foggy weather, higher in sunny weather, and highest in overcast weather. All expectations were derived from Vinod (2007;personal communication, 2018), Robin et al. (2014) and preliminary field surveys.  (2012)).

215
We tested 127 detection structures with an occupancy structure that included three covariates and 216 used the best-fitting detection structures to fit 56 occupancy covariate structures. The best-fitting 217 occupancy structures (AIC weight >0.02) are in  (Table 3). Seven of the 11 independent variables appeared in structures with 233 substantial support. We used log(site area) as an offset in all N-mixture models to control for site 234 area, so that estimates of site-specific abundance λ could be interpreted as expected Nilgiri pipit 235 density per hectare.

Nilgiri pipit abundance 259
In the 144 sites with ψ > 0.4, we detected 0 to 14 individuals (mean = 3.76). Despite excluding 260 sites with a low probability of occurrence, the χ 2 goodness-of-fit test based on 5,000 parametric abundances relative to the model (=1.93). Estimated ̂ was therefore used to derive QAIC 263 values for model selection, and to adjust estimated variances. Model-averaged results showed 264 that elevation had a strongly positive effect on bird density: the predicted density at the 265 maximum sampled elevation was more than twice that at the lowest elevation. The best-fitting 266 six models all included elevation, grass height, wattle maturity, and water, while Rhododendron, 267 Eucalyptus, and recent burns appeared in three of the four best-fitting models. Eucalyptus size and isolation, and plantation extent, were not part of models that had any substantial support.

Discussion
We found that maximum site elevation, grassland patch size, and distance to the nearest large grassland were the only covariates that had substantial effects on pipit occupancy. In contrast, abundance was shaped by maximum site elevation in combination with many site-level habitat characteristics, each of which had a substantial effect on predicted abundance.
Nilgiri Pipit occupancy and abundance both have strong relationships with elevation. Only sites above 1800m have high probabilities of occupancy. Thus, the Nilgiri pipit is likely to be the most elevationally restricted bird species in the Indian subcontinent south of the Himalaya. The only detection of Nilgiri pipits below 1700m was at the southwestern extremity of the Anamalai plateau, with anomalously high exposure to the southwest monsoon and microclimate consistent with much higher areas elsewhere. This relationship between occurrence and elevation is consistent with previous research indicating that montane grasslands become the dominant biome only above 2000m (Das et al., 2015), and is similar to the distribution patterns of other montane flora including Rhododendron (Giriraj et al., 2008) and fauna (Yandow, Chalfoun, & Doak, 2015;Mizel, Schmidt, Mcintyre, & Roland, 2016) including habitat-specialist birds (Watson, 2003) like the Sholicola (Robin & Sukumar, 2002). This sensitivity to elevation makes the Nilgiri Pipit extremely vulnerable to climate change: it is likely to undergo substantial range contraction due to anthropogenic global warming (Parmesan, 2006;Sekercioglu, Schneider, Fay, & Loarie, 2008), given the distribution of area with respect to elevation in the two plateaux (Table 4). Any loss of habitat for restricted-range species such as the Nilgiri pipit can be severely detrimental. Our sampling did not include patches of dense, mature, plantations of the invasive woody species, as these habitats are not viable for grassland specialists. Within the sampled grasslands, the presence of these invasives did not affect the occupancy of the Nilgiri pipit, but negatively affected abundance. The exception was the patch condition "immature wattle", which had a higher predicted abundance than either "mature wattle" or "no wattle". We suggest two possible explanations for this anomaly. First, areas with immature wattle but without mature wattle largely (19 of 28 sites) occur within the two largest grassland patches, Eravikulam and Mukurthi National Parks, which have high Nilgiri pipit densities, and where management practices include removal of mature wattle. Second, immature wattle may temporarily contribute to habitat heterogeneity (a factor that increases abundance, as discussed below) without substantially degrading habitat quality. However, due to wattle's rapid growth, this effect is likely to be transient.
Our findings suggest that while Nilgiri pipit presence may be constrained by habitat availability, suggesting that such a spread is a widespread phenomenon globally, requiring broader attention.
It is probable that low-density populations in areas affected by invasive species are non-viable.
In grasslands that are remnants of century-old habitat loss in the eastern Nilgiris (Joshi et al., 2018) we found a complete absence of Nilgiri pipits, but grassland remnants in the Palani Hills, invaded and fragmented severely since 1973 (Arasumani et al., 2018) supported low-density populations. We conclude that the eastern Nilgiris have experienced local extinction, as historical records of the species exist from that region (Robin et al., 2014). Furthermore, high-altitude habitat specialists are often strongly affected by patch area (Watson, 2003), but the Nilgiri pipit showed only a weak correlation with patch size and isolation, which is unexpected for a poor disperser (Rosenzweig, 1995), such as the Nilgiri pipit (Vinod, U., personal communication, 2018). These findings may represent a substantial extinct debt in grassland recently affected by invasive species in the Palani hills and the southern Anamalais. The effects of local extinctions on population structure and viability is likely to be stronger in non-avian grassland endemics, since birds have greater dispersal abilities (Watson, 2003), particularly for long-lived species that are likely to have greater extinction debt (Krauss et al., 2010).
The abundance of the Nilgiri pipit showed a strong positive correlation with intermediate or mixed grass height: such a preference for specific grass height has also been documented for a suite of avian species in the Brazilian pampas (Jacoboski et al., 2017). We found moderate positive correlations with other variables contributing to local habitat heterogeneity and vegetation structure, a correlation supported by the behavioural observations of Vinod (2007), who found that Nilgiri pipits preferred marshy habitat with tall grass for nesting, and more open habitat for feeding. The relationship between habitat heterogeneity and grassland birds is complex, with both positive and negative responses documented (Wiens & Rotenberry, 1981;Pavlacky, Possingham, & Goldizen, 2015). While further investigation of the functional effects of structural complexity in the montane grasslands of the Western Ghats is merited, our findings show a broad dependence on natural habitat heterogeneity within this habitat. We found that patch-level habitat quality had a strong effect on abundance: such a pattern has also been found in other highland avifauna (Allan et al., 1997;Watson, 2003). Thus, conservation efforts must focus on maintaining habitat quality over and above simply preserving grassland. Frequent fire is thought to positively affect grassland avian species richness (Pons, Lambert, Rigolot, & Prodon, 2003). We did not find a strong relationship between recent burns and Nilgiri pipit density: further systematic study is required to draw any conclusions about the role of fire in management in this landscape. At a local scale, the presence of woody invasives reduces Nilgiri pipit abundance, while at the landscape scale, the spread of the same woody invasives shapes grassland patch size and isolation (Arasumani et al., 2019), which drive Nilgiri pipit occupancy. We emphasise that our study was limited to the remnant grasslands, and our findings therefore underestimate the detrimental effects of invasive vegetation, as completely wooded habitats that were grasslands earlier were not sampled since these do not have any grasslands or pipits. Controlling the spread of invasive tree species is a matter of urgent conservation attention.

Conclusions
We demonstrate that elevation shapes both occupancy and abundance of this montane specialist.
Furthermore, our study demonstrates that woody invasives are constraining occupancy via rapid grassland loss at the landscape level, while also degrading habitat quality at the local level. Our research also indicates local extinctions in large parts of the species' range, and probable extinction debt in other parts. This study underscores the urgent need for conservation actions targeted at the poorly known montane grasslands and the specialist species dependent upon it. this manuscript were made after AL started graduate school at the University of Chicago, and we thank them for their support. This research was carried in compliance with all relevant institutional norms and Indian laws.

Author contributions
VVR, CKV, DJ, and AL conceptualized and designed the study. AM collated and analysed remotely sensed data. AL and CKV collected field data. AL analysed field data with help from DJ and VVR. AL wrote the manuscript with input from all authors.

Conflict of interest
The authors declare that they have no conflict of interest.

Administrative regions examined in the study
Our study area fell within the following administrative regions. ). The accuracy of the classified map was calculated using ground truth GPS points. We created 100 random points on the map and visited each of them to evaluate the accuracy of our classification. Overall accuracy was found to be 96.5%, and the kappa coefficient was 0.93 (following Congalton, 1991).

Sampling sites and site selection
Of the 3012 grassland patches which remained after small patches had been removed from our sampling frame (sensu Williams, Nichols, & Conroy, (2002)), 300 were randomly selected to be surveyed for detection/non-detection of Nilgiri pipits. Of these 300, areas containing 69 sites were removed because they were determined by the relevant government agency to be unsafe due to armed militant activity, or for other reasons of accessibility. The remaining study landscape included 234 sites. Of these 234, 8 sites were removed due to misclassification: they were found to be Eucalyptus plantation or scrub forest, and in 3 cases, to be under water, because they occurred on the banks of a reservoir in which the water level had risen between the point at which classification was carried out, and the time at which the surveys were conducted. A further 21 sites could not be sampled as they were directly adjacent to other (sampled) sites, and Nilgiri pipit presence in them could not be assumed to be independent of all other sites. Thus, the total area of grassland sampled was 434.98km 2 , which contained 202 sites that fit our criteria for sampling (accurate habitat classification, and independence from other sites. Of these 202, 170, or 84%, were sampled: 32 sites were found to be physically inaccessible due to the topography of the landscape or the ownership of the land.

Covariates
Landscape-level covariates were generated from GIS data. Site-level covariates were generated either from GIS data or field observations. Visit-level covariates were measured in the field. A full list of covariates is provided below. Variables used in the final analyses are in boldface font.
Of the 22 site-level independent variables, we eliminated some variables based on collinearity.
Five variables were eliminated because there was insufficient variation in these across our sites.
Any other covariates displaying moderate collinearity were not included within the same model.
Of the 11 remaining covariates, two were expected to affect abundance but not occupancy, and some were used only as detection covariates in modelling pipit occupancy. mearnsii were divided into the categories "Mature" and "Immature", as each category was expected a priori to affect pipits differently and these also seemed to accurately reflect variation on the ground, ; several sites had experienced recent invasion from A. mearnsii and had not yet been substantially altered by this invasion. Furthermore, black wattle trees experience die-offs beyond a certain size, which are easily identifiable in the field, providing a good proxy for the maturity and hence density of a stand of wattle.

Supplementary
The threshold of 1.5km 2 for defining a "large grassland" was based on observations which found that virtually all pipit populations which appeared large and healthy, and where pipits were regularly detected in large numbers, were in patches larger than 1.5km 2 .
The predominant height of the grass within a site was a categorical variable with three level.
Grass was characterized as short when it was shorter than 15cm, tall when it was taller than 75cm, and intermediate in all other cases. This categorization was based on observations of different microhabitats. Cliffs, heavily grazed areas, or recently burned areas had extremely short grass; marshy areas and scrub habitat found up to approximately 1800m had very tall grass; the 'intermediate' category typically included native montane grasslands and more mixed grassland.
Visit-level covariates included the weather, date, time of day, and identity of the most experienced observer. Weather was recorded as a categorical variable with three levels; sunny, overcast, and foggy, as these types of weather had clear effects on detection. Data were not collected during rain. Observer identity was recorded to assess potential differences in detection efficacy between observers. The number of days since the first survey was recorded as a proxy for season, as seasonal variation in the Nilgiri pipit's behaviour was known (Vinod 2007) and needed to be controlled for. Similarly, the diurnal activity of the Nilgiri pipit was known to have a bimodal pattern with peaks soon after sunrise and shortly before sunset: thus, we transformed time of day to time away from solar noon for analysis.
Gorse, Scotch Broom, Pine, and Lantana were eliminated as variables because they were found to be present in <10% of the sites surveyed. Eupatorium and the "Other native shrubs" variable were eliminated because they were found in >95% of the sites surveyed.