Occupancy of the Ethiopian endemic Moorland Francolin in pristine and degraded Afroalpine biome using a camera trap approach

Abstract Occupancy modeling is an essential tool for understanding species‐habitat associations, thereby helping to plan the conservation of rare and threatened wildlife species. The conservation status and ecology of several avian species, particularly ground‐dwelling birds, are poorly known in Ethiopia. We used camera trap‐based occupancy modeling to investigate habitat covariate influence on occupancy (Ψ) and detection probability (ρ) estimates of Moorland Francolins Scleroptila psilolaema from spatially replicated surveys across both relatively pristine and disturbed landscapes in the Afroalpine biome of Ethiopia. Model‐averaged estimate of ψ^ across all sites was 0.76 (SD = 0.28) and ρ^ was 0.77 (SD = 0.13) in the pristine landscape. The ψ^ of the species in the disturbed landscape was 0.56 (SD = 0.19) and ρ^ was 0.48 (SD = 0.06). As hypothesized, based on our model‐averaged beta coefficient estimates (βmean ± SE), predators significantly negatively influenced the occupancy of Moorland Francolins in pristine habitat. We also found a significant positive association of occupancy with herb species richness. Contrary to our prediction, distance to road significantly negatively influence the occupancy of the species, suggesting that occupancy probability was highest in proximity to roadsides and trails in the pristine habitat. There was no significant influence of habitat covariates on the occupancy of the species in the disturbed habitat. The most important covariates that significantly influence the detectability of the species in pristine habitat included sampling occasion and precipitation. The greater occupancy and detectability of this endemic species in the pristine habitat could be linked with the particular conservation status and management of this biodiversity hotspot in the central highlands of Ethiopia. Our results suggest that strict legal enforcement is required to sustainably preserve Moorland Francolins and the ecological integrity of the entire Afroalpine biome. We recommend using camera traps in order to develop realistic and effective conservation and management strategies for rare, sensitive, cryptic, and ground‐dwelling animals in the region.


| INTRODUC TI ON
Among the 34 Earth's biodiversity hotspots, the Eastern Afromontane hotspot, including the Ethiopian Highlands, ranks fourth by a number of endemic plant and vertebrate families and genera (Mittermeier et al., 2004).Next to the Guinea-Congo Forests biome, the second-highest number of biome-confined bird species are found in the Afrotropical Highlands biome (BirdLife International, 2004).In Ethiopia, all bird species subsist in three biomes: the Afrotropical Highlands (including the Afroalpine and Afromontane), Sudan-Guinea savanna, and Somali-Masai biomes (Fishpool & Evans, 2001;Gedeon, Zewdie, & Töpfer, 2017).The Afroalpine biome of Ethiopia consists of a complex mosaic of grassland, moorland, bushland, and other habitat types which are unique in terms of species distinctiveness.This biome harbors a considerable endemic flora and fauna and is home to a number of range-restricted bird species (Ash & Atkins, 2009;Gedeon, Zewdie, & Töpfer, 2017;Töpfer & Gedeon, 2020), as well as to rodents (Ashenafi et al., 2012;Bryja et al., 2019;Razgour et al., 2021), and medium and large-sized mammals (Ashenafi & Leader-Williams, 2005).
Like in other tropical countries, the distribution of vegetation in Ethiopia reflects the interplay among altitudinal variation as well as climatic and other abiotic factors (Friis et al., 2010).The combination of different habitat characteristics, species traits, and their interactions define the occurrence, occupancy, and abundance of wildlife populations and influence their distribution patterns and detectability (Devarajan et al., 2020;Guillera-Arroita, 2017).
Knowledge on its breeding biology, home range size, population abundance, occupancy (i.e., habitat use), and other ecological patterns is still scant.Previous distributional data showed Moorland Francolins to occur in the eastern and western highlands (Ash & Atkins, 2009;Gedeon, Zewdie, & Töpfer, 2017).It is classified as Near Threatened due to the ever-increasing loss of moorland and grassland habitats (BirdLife International, 2023), but its population size and habitat association along its geographical range are insufficiently known.
We attempt to draw an inference of baseline data on the ecology of Moorland Francolins using an occupancy modeling framework.
We used presence/absence (i.e., detection/non-detection) data to analyze two stochastic processes: occupancy and detection probability.Occupancy is a dichotomous state variable that accounts for imperfect detection to minimize unreliable inferences of species distribution and range (Bailey et al., 2014;Guillera-Arroita & Lahoz-Monfort, 2012;Kéry et al., 2010;MacKenzie et al., 2018;Tyre et al., 2003).Occupancy models estimate the probability of a species' presence in a fraction of landscape units (MacKenzie et al., 2002(MacKenzie et al., , 2018) ) and help to understand habitat use within a landscape.They are applied across several animal taxa for the implementation of successful conservation and management strategies (Burton et al., 2015;MacKenzie et al., 2018;Steenweg et al., 2017).
Therefore, the objective of this study was to gain insight into the habitat use of Moorland Francolins in its native range for the first time and to investigate the effect of habitat covariates on occupancy and detection probability from spatially replicated surveys.

| Study area
This study was performed in two areas (Figure 2): Guassa Community Conservation Area (hereafter GCCA) and an area encompassing Sululta plain, Entoto Natural Park, Ankober-Debresina escarpment, and a few sites between them (hereafter collectively abbreviated SEA).The study areas are part of Ethiopia's central highlands in which several Important Bird and Biodiversity Areas (IBAs) are designated (Tilahun et al., 1996).These highland areas consist of top mountain massifs and volcanic cones (Friis et al., 2010).Most of our study sites (93%) were located in IBAs, including GCCA, Entoto Natural Park, Ankober-Debresina escarpment, and Sululta plain.The remaining sites were located outside these IBAs in Angolela Tera, Assagirt, Sheno, and Mendida districts.However, both IBAs and non-IBAs sites in SEA are under serious anthropogenic threat: farming, livestock grazing, settlement, monocultural plantations, and recreational activities.For instance, ENP has shifted its purpose from conservation implementation (Tilahun et al., 1996) to recreational area where mass tourism (Asefa, 2018;Tesema & Berhan, 2019) and monocultural plantations (Bahru et al., 2021;Tadesse & Tafere, 2017) strongly affect the landscape.Both Sululta plain and Ankober-Debresina escarpment are mainly influenced by livestock grazing, farming, and settlement expansions.Except for Sululta plain, the other areas are dominated by exotic Eucalyptus plantation and African juniper Juniperus procera (Esayas & Bekele, 2011).Therefore, we distinguished between the two study areas based on their different levels of human disturbance, topography, floristic structure and composition, and conservation status, considering GCCA a relatively pristine and SEA a strongly humanmodified area. 2 and 3) covers 78 km 2 (Steger et al., 2020), yet the total land area sums up to 111 km 2 if the adjoining villages and other land use types are included (Ashenafi & Leader-Williams, 2005;Nigussie et al., 2019).This area shows critically important habitat features for many wildlife species (Steger et al., 2020) and comprises both the Ericaceous belt (3000-3200 m a.s.l) and the Afroalpine belt (above 3200 m a.s.l) (Friis et al., 2010).The area has been managed by the local community through a management model called the Qero system (Ashenafi et al., 2012;Ashenafi & Leader-Williams, 2005) for the conservation of the species itself (Eshete et al., 2015;Estifanos et al., 2018). 2 and 3) forms part of the Afromontane with altitudes generally below 3000 m a.s.l.Very small patches of herbs, shrubs, scattered acacia trees, and exotic trees are common.Here, the Moorland Francolins persist in very small uncultivated and grassland patches of Afromontane habitats.

SEA (Figures
These highland areas experience a bimodal rainfall pattern with main rain from June to September and smaller amount of rain from October to February (Mohammed et al., 2022).The distinctive habitat features of both of these areas are erratic climatic conditions and a very short dry season (ca. 2 months).The mean annual temperature of GCCA and SEA are 21.26°C (± 0.95 SE) and 15.53°C (± 0.55 SE), whereas the mean annual precipitation of GCCA and SEA were 2.65 mm (± 0.78 SD) and 2.69 mm (± 0.90 SD), respectively (Figure S1).

| Sampling design
Site selection for this study was made randomly.Most sites were obtained through a distribution map from the IUCN, scientific literature, and citizen science data, whereas some sites were chosen without antecedent species records.Following the standard design procedure for allocating optimal sampling occasion (MacKenzie (n = 20 for GGCA and n = 19 for SEA), with an average transect length of 2.04 km (± 0.80 SD) across both study areas.In this study, we expected that the number of sites (s) and occasions (K) were sufficient to determine the stochastic processes.Then, the total survey is simply defined as s × K, and the maximum survey occasion for each site was calculated by minimizing s, while taking a standard error of 0.05 for GCCA and 0.065 for SEA.Since both study areas are separated by approximately 150 km independent camera trap data collections were conducted for 5 months for both areas.Along these geographical scales, specific habitat characteristics (i.e., covariates) predicted to influence the occupancy and detection rates of the target species were measured at each site (Table 1).

| Camera trapping
In December 2019 and the first 3 weeks of January 2020, we made a pilot survey in both areas to assess the study species using camera traps and broadcast playback methods.A total of 20 cameras (Browning Trail Cameras and Bushnell Trophy Cam HD brands) were used for short-term deployments in this study.Since we had a small F I G U R E 3 Afroalpine habitats in the central highlands of Ethiopia: GCCA with most habitat types (a) and the target species feeding in Helichrysum-Festuca grassland (c) and SEA with degraded rocky habitat (b) and grazing land (d).(Continues) number of cameras, some adjoining habitats (see habitat covariates below) were simultaneously assessed and in both study areas cameras were deployed sequentially.Cameras were repositioned to other sites to cover the desired representative home range and to make the field survey more cost-effective.When small camera traps are available, repositioning to new sites is recommended to increase the spatial coverage of target species (Meek et al., 2014;Shannon et al., 2014;Si et al., 2014;Wearn & Glover-Kapfer, 2017).
Each camera trap was placed horizontally (i.e., camera alignment was perpendicular to the ground) within a 50 m radius (~ 0.8 ha) of plot or focal patch size to optimize detectability.Because some terrain settings were very difficult to conduct surveys, cameras were not fixed at the center of each plot instead they were placed approximately 10-30 m distance from the grid center, where freshly raked and possible feeding grounds were noticed.Single camera placement is employed to detect small-medium mammals and bird species (Ferreguetti et al., 2015;Lamelas-López & Salgado, 2021).
The camera spacing in continuous habitats in GCCA was approximately 0.3 km (0.2-0.5 km), while in SEA was approximately 0.5 km (0.3-0.8 km) to enhance detectability and to avoid spatial autocorrelation between camera traps.Though telemetry data collection was originally proposed to estimate the home range of the species which enables to estimate camera spacing, we assumed that the camera trapping space was sufficient and representative to study occupancy of this species based on available literature.If the average home range size of a target species is not known, it is recommended to infer spatial extent from congeneric or other related species (Niedballa et al., 2015).Mostly, camera trap spacing, based on home range, for pheasants ranges from 0.2 km (Zou et al., 2019) to 0.7 km (Suwanrat et al., 2015).Therefore, the camera spacing was higher than the home range diameter of the species, which was a similar approach as in other studies (Maffei & Noss, 2008;Niedballa et al., 2015).In our case, camera traps were unbaited but rather were providentially camouflaged with rocks, stones, and Ericaceous heathlands of the study sites.Site selection for camera placement was randomly carried out across various habitats of both study areas, as was proposed by several other studies (e.g., Burton et al., 2015;Cordier et al., 2022;Meek et al., 2014;Tanwar et al., 2021;Wearn & Glover-Kapfer, 2017).
We placed camera traps on tree trunks, attached to thick coarser grasses (Festuca spp.) and shrubs, and on wooden stakes at approximately 30-60 cm above the ground, as this standard height is credible to trigger the motion sensor and it is reasonable to detect ground-dwelling bird species (Figure 3; Figure S2).Because some sites were in completely rocky areas, we also put cameras by arranging stacked stones that matched the background of the site.Most cameras had 16 GB memory and some cameras mounted on courser grasses and shrubs had 32 GB SanDisk memory card as they were easily triggered by the movement of vegetation during high wind velocity.However, to enhance good photographs and detectability, prudent vegetation removal was carried out in some sites to avoid false triggering mainly during windy conditions (Meek et al., 2014;Wearn & Glover-Kapfer, 2017).Our primary interest was to capture photos of the target species that can be easily pooled into detection/non-detection binary matrices.In most cases, the video function was discounted, yet some videos were collected from the field Note: The first nine predictors are site-specific covariates, whereas the last four are observational-specific covariates. a The spatiotemporal covariates are dropped due to high collinearity (Dormann et al., 2013;Zuur et al., 2010).This study selected herb species richness over total species richness (Sprich) in both study area.Herbaceous and shrubby vegetation were dominant in GCCA (> 80% ground vegetation cover) (Nigussie et al., 2019).b Hunting was not considered as a threat for this species (See Discussion).
c Human disturbance factors: grazing, mowing and farming are the major factors in the study sites (Ashenafi et al., 2012;Nigussie et al., 2019;Steger et al., 2020).Festuca abyssinica grass (Guassa) intriguingly is valued for fodder for livestock (cut and carrying system and livestock grazing), thatching, wall building mix with mud, and help to make whip, rope, hat, broom (mure) and raincoats (gesa).

TA B L E 1 (Continued)
to understand the natural behavioral repertoire of the species and its interaction with other species (i.e., predators) in the habitats.Because both camera models had different setting options but similar functions, we set up cameras for the following typical important parameters: (1) camera traps were active for 24 h/day and programmed to capture 1 photo/trigger at 10 s intervals, and some sites with more than one camera traps set to capture 20 s video/trigger, with subsequent videos delayed for 5 min; (2) the sensitivity of the infrared sensor was programmed to be medium or normal; and (3) the quality of photos were adjusted to be medium for both camera brands.
The battery life of each camera was checked during data retrieval, storage, and repositioning of cameras.Extreme weather conditions (too hot or too cold) severely affected the sensitivity of sensors in our areas.

| Habitat covariates
To include representative habitat types in GCCA, we adapted the habitat classifications of Ashenafi et al. (2012).The habitat types were Mima Mound, Erica Moorland, Euryops-Alchemilla shrubland, Helichrysum-Festuca grassland, and Festuca (Guassa) Grassland.In their classifications, swamp habitat which is typically characterized by woody vegetation (US definition) and reed swamp or forested fen (European definition) is now replaced by "peatland".In this habitat, the wetland type is normally a moor surrounded by Erica, Festuca and other plant species and has permanent and ephemeral water fed by precipitation hence called "ombrothropic peatland".Moreover, we identified and added montane forest to the classification as an important other habitat type for wildlife species in the area, though it was not included in the rodent-based study (Ashenafi et al., 2012).
Because the sites in SEA study area were human-dominated, the habitat types were homogenous and it was very hard to distinguish and classify in relation to vegetation patterns.Broadly, we categorized the habitats into Eucalyptus-Juniperus habitat and grazing lands.
The later class obviously incorporated agricultural lands.Overall, this area has been heavily transformed to Eucalyptus plantations to meet demand for wood products and improve the livelihoods of local communities (Bahru et al., 2021;Tadesse & Tafere, 2017).
At the sites, we collected 13 covariates derived from habitat features, landscape connectivity metrics, climatic factors, and sampling covariates which were predicted to influence the occupancy and detection probabilities of the target species.Occupancy was modeled as a function of site-specific covariates, including biotic factors (vegetation traits and predators) and landscape connectivity metrics, while detectability was modeled as a function of observationalspecific covariates, including survey occasion (hereafter occasion) and climatic factors (precipitation and temperature).The occasion is defined as a total number of days for which each camera was active per site (Table 1).
Specific vegetation traits assumed to influence habitat use were collected from each site using different tools.Due to the occurrence of scattered trees within most sites (with the exception of montane forest adjoining to the moorland habitats and ENP) and complex landscapes varying with soil, climate, topographic, and other features, we used only two 20 × 20 m 2 randomly placed quadrats for tree species with DBH ≥ 10 cm in woody vegetation sites separated by at least 15 m between quadrat.Meanwhile, in each large quadrat, 5 × 5 m 2 for shrub and liana species with ≤ 10 cm were nested (Figure S2).Thus, the following vegetation traits were measured accordingly: (1) by placing five 1 × 1 m 2 quadrats (four in the corner and one in the center) in each nested quadrat; herb and fern species richness was identified and counted; (2) woody species richness and abundance were determined from the larger and nested plots; (3) woody species density (abundance of individual trees, shrubs and lianas/0.8ha) was also estimated from each site; and (4) average tree canopy cover was estimated using GLAMA (Gap Light Analysis Mobile Application software) from vertically upward looking photos (approximately 8 photos/site) either directly collected in the field or retrieved photographs with a digital camera (Nikon D5300) from sampling sites (Gonsamo et al., 2011;Tichý, 2016).
Landscape connectivity metrics (landscape scale covariates), including elevation, distance to the nearest road (both paved and unpaved roads and trail with at least 1 m wide), distance to nearest water points, and distance to nearest settlements were gauged either directly at the site using a handheld GPS and tape meter or indirectly using Google earth images.Nearest and accessible metrics to some sites were measured in the field.Average on-site ambient temperature and precipitation measurements would have been costly and very difficult to conduct in each site; instead, we obtained climatic data from NASA 2022 (https://power.larc.nasa.gov/data-access-viewe r/) to understand species-habitat associations.

| Data analysis
Single-season occupancy model was applied to understand the influence that habitat covariates have on occupancy and detectability while accounting for imperfect detection (MacKenzie et al., 2002(MacKenzie et al., , 2018)).The detection history was derived from a sequence of species detection/non-detection dichotomous data (i.e., detection = 1 and non-detection = 0) that were pooled into occasions from consecutive camera days for each site.For occupancy models, data collected by camera traps needs to be divided into sampling occasions (Sollmann, 2018).Such data treatment is important to maximize detectability, maintain spatiotemporal independence among occasions and thereby increases adequacy of model fit.Sensitive analysis was conducted without incorporating any covariates to evaluate the discrepancy of occupancy and detection estimates for different sampling intervals.Based on the input of the analysis, we chose the balance between high parameter estimates and small confidence intervals (see Table S1).Consequently, an occasion was defined as an interval of two camera days for both study areas.
Cameras were active for approximately six consecutive days This also included data from malfunctioned cameras and blank photos in some cameras.
We used PRESENCE program v.2.13.39 (Hines, 2006) to model occupancy and detection estimates.The parameters were estimated using logit link and a maximum likelihood approach in the program (MacKenzie et al., 2002(MacKenzie et al., , 2018)).Occupancy probability (Ψ) was modeled as a logit link function of fine-scale level and landscape scale covariates.The structure of model framework of the occupancy probability of a site (i) in association with the site-specific covariates is expressed as: Likewise, the detection probability (ρ) was modeled as a logit link function of observation-specific covariates.The logit equation derived from the probability of detecting a species at site i, during survey j in association with the covariates is: where X i1 … X iu refers to site covariates associated with the probability of a site i being occupied and y ij1 … y ijv refers to sample covariates.
All continuous covariates were normalized by z score conversion (mean = 0 and SD = 1) to help convergence of the maximum likelihood algorithm prior to analysis (Schielzeth, 2010).Such data transformation produces better model performance and interpretability (Gelman & Hill, 2007;Schielzeth, 2010).Since we had spatial data, collinearity was assessed using variance inflation factor (VIF).Covariates with highest VIF were dropped in the analysis and covariates at threshold level VIF < 5 and Spearman's correlation (r s < 0.7) were retained (Dormann et al., 2013;Zuur et al., 2010).Of the strongly correlated covariates, we retained ecologically important covariates based on field evidence and existed literature to understand their influence on occupancy and detectability.With a total of 11 covariates, the global model was run, and subsequently competing models were constructed based on plausible additive covariates.The null model (ψ(.),ρ(.))was also constructed to compare with the relative weight of other additive models which included one or more covariates.
Since the ratio of effective sample size to the number of parameters (n/k) was small, model selection procedures were carried out using Akaike's Information Criterion for small sample bias adjustment (AIC c ) from the competing candidate set of models (Burnham & Anderson, 2002), where the most supported models are top-ranked models with ΔAIC c ≤ 2.0 (Burnham & Anderson, 2002).
Summed model weights of each covariate from each model were also calculated to rank the relative importance of the covariates (Burnham & Anderson, 2002).Then, in order to retain ecologically meaningful covariates, models with ΔAIC c ≤ 4.0 were selected to drive model average estimates of occupancy and detection probabilities (Burnham et al., 2011) (Tables 2 and 3).Competitive models were used to estimate Ψ and ρ and calculated parameter estimates, standard errors (SEs), and level of significance based on 95% CI (zero-overlapped method) for each covariate.Uninformative parameters (Arnold, 2010;Leroux, 2019) were also assessed in our model sets.Estimates of the slopes (i.e., β coefficients) for covariates were used to determine the magnitude of their influence on Ψ and ρ.
We used a parametric bootstrap goodness of fit (GOF) using We computed the number of occasions (K) to enhance the odds of detecting Moorland Francolins in a site.We considered a set of four levels of confidence (ρ*): 0.7, 0.8, 0.9, and 0.99 by assuming that the species detection probability is always less than one.The occasion (K) was calculated from the detection probability (ρ) of the model averaging to determine the true absence of the species from a site (McGrath et al., 2015;Pellet & Schmidt, 2005;Sewell et al., 2010).
where ρ is the calculated detection probability and ρ* is the target detection probability as mentioned above.
Both original and square-root transformed data were used sequentially for normality assumption using Shapiro-Wilk and homoscedasticity tests.Consequently, we used one-way ANOVA to compare mean differences in photos captured among sampling months in GCCA, and a post hoc testing procedure using Bonferroni error adjustment was applied for multiple comparisons.We also used Mann-Whitney U-test to compare mean differences in photos and parameter estimates across spatiotemporal.Similarly, this test was used for occupancy probability estimates comparison in relation to predator presence and absence.This data was analyzed in

IBM SPSS statistics (version 20). A two-tailed hypothesis test with
an alpha value of 5% was considered.

| Camera trapping in GCCA and SEA
The deployed camera traps yielded 610 and 361 trap nights in GCCA and SEA, respectively.We failed to collect data from 21 (GCCA) and SEA (18) sites mostly due to battery failure and system malfunctioning.We found a significant difference in average photos captured among sampling months in GCCA (F 2,95 = 11.775,p < .001).There was no significant difference in average photos captured between sampling months in SEA (Mann-Whitney test U = 277.5,n = 48, p = .893).
Pooling the data across both study areas, the average photos captured in GCCA was approximately four units higher in comparison to SEA (Mann-Whitney test U = 1365, n = 146, p < .001)(Figure S3).

Likewise, model-averaged estimates of occupancy probability ( ψ)
and detection probability ( ρ) parameters were significantly higher in the pristine habitat than in the disturbed landscape (Figure 4).

| Habitat use modeling for traditionally managed habitat
We captured a total of 2632 photos (7-141 photos per site) from all sampling occasions in GCCA.sites) estimate of 0.69.In GCCA, at the habitat-specific level, the findings showed that the highest habitat use was obtained in Mima Mound, Euryops-Alchemilla shrubland, and Helichrysum-Festuca grassland.Conversely, the least was shown across the tree belt (i.e., montane forest and Eucalyptus plantation) (Figure S4).
The null model (ψ(.), ρ(.)) appeared to be the least important model to explain the stochastic processes (Table 2; Table S2).The Ψ for this model was 0.73 (SE = 0.05) with a 95% CI of 0.63-0.82and ρ of 0.85 (SE = 0.03) with 95% CI of 0.79-0.89.In GCCA, some evidence of breeding activity was observed from the camera traps, such as three juveniles were provisioned by both parents.
We constructed candidate sets without interactions between covariates to model Ψ and р in the order of parsimony models using   showed model selection uncertainty (Symonds & Moussalli, 2011).
Due to the ecological importance of individual covariates included in the top models, we discounted models with less than five ΔAIC c to increase model weight (Richards, 2005) and we considered the top-ranked models with summed model weight of 0.95 (Symonds & Moussalli, 2011).
These predators were avian and mammalian species.We also found that herb species richness showed a significantly positive influence the occupancy of the species based on model averaging estimates (β mean ± SE = 1.40 ± 0.68, 95% CI: 0.07-2.74)and the summed ω i was 97% (Table 3; Figure 6).Contrary to our prediction, distance to road was significantly negatively influenced the Ψ of the species and the model weight of the covariate was 78% (β mean ± SE = −0.74± 0.35; 95% CI: −1.44, −0.05), suggesting that occupancy probability decreased as the distance to road increased in the pristine habitat (Table 3; Figure 6).
As depicted in the top models, the ability to detect Moorland Francolins was modeled as a function of survey occasion, precipitation, and temperature with summed model weight of 0.95, 0.92, and 0.70, respectively.The most important covariates supported by our hypotheses, however, included sampling occasion (β mean ± SE = 0.68 ± 0.23, 95% CI: 0.23-1.13)and precipitation (β mean ± SE = 0.75 ± 0.36, 95% CI: 0.05-1.45),both of which significantly positively influenced the detectability of the species (Table 3; Figure 6).Although the detectability of the species was increasing with temperature, the beta coefficient estimate (β mean ± SE = 0.40 ± 0.23; 95% CI: −0.04 to 0.84) overlapped zero which exhibited a positive association but non-significant difference with habitat use of the species.

| Habitat use modeling for human-modified landscape
In the human-modified landscape, a total of 339 photos (2-29 photos per site) from 23 sites were trapped, yielding a naïve occupancy estimate of 0.48.The Ψ estimate without any covariate was 0.54 (SE = 0.08) with a 95% CI of 0.38-0.70 and ρ of 0.54 (SE = 0.06) with a 95% CI of 0.42-0.65.In this study area, based on the above considerations, the null model was included in the top important models with ω i = 0.95 to explain the stochastic processes.The global model (ψ(WD + Hsp + T caco + Pre + Elev + DR + DS + DW),ρ(E + T + P )) showed no evidence of lack of fit (χ 2 = 118.13;p = .35;ĉ = 1.07).
Model-averaged estimate of ψ across all sites in SEA was 0.56 (SD = 0.19) and ρ was 0.48 (SD = 0.06).The overall occupancy was underestimated by approximately 17% when detection probability is not accounted for.Distance to settlement, tree canopy cover, herb species richness, distance to road, predator, and woody density appeared in the competing models to explain habitat use of the target species in this area.As predicted, distance to settlement (ω i = 0.76; β mean ± SE = 0.74 ± 0.41; 95% CI: −0.07 to 1.55) positively associated with habitat use of the species, yet its respective 95% CIs slightly overlapped zero.Other covariates also showed non-significant associations with occupancy of the species (Table 3).In this study area, detectability was more supported without covariates based on the top models.Thus, the sample covariates predicted to influence detectability had relatively low summed weight and 95% of CIs overlapped zero.In this disturbed habitat, detection probability was not significantly affected by sample covariates but all covariates depicted positive association with detectability.The summed model weight of each covariate was below 0.30 (Table 3).

| Recommended number of sampling occasions (K)
The sampling occasion (K) needed at GCCA was ranged from 1 to 3, this meaning that a single occasion (mean 0.86 and 1.14, respectively) was needed for a targeted confidence level of probabilities of 0.7 and 0.8 and two (mean 1.64) and three (mean 3.27) occasions sequentially were sufficient for 0.9 and 0.99 detection probabilities to estimate the true absence of the species at a given site.Similarly, we found that 2, 3, 4, and 7 occasions sequentially were needed at SEA.

| Occupancy and detection probability estimates using camera trap
Our study delivers the first insights into the habitat use of Moorland Francolins using a camera trap approach.Camera traps for this elusive and cryptic species helped us to avoid false-positive detection, which also corroborates the respective assumption for the occupancy model.The overall or true occupancy estimates in both study areas were greater than the naïve occupancy (ψ) estimates when detection probability is accounted for.These suggest that models incorporate imperfect detections to discount underestimating In many tropical African countries, the protected areas are called "paper parks"-existing in name only as they poorly counter habitat and species loss (Dudley & Stolton, 1999).However, GCCA as a traditionally protected area is exceptional in this case as the indigenous knowledge for conservation of natural resources, the Qero system, has supported several wildlife species for almost four centuries (Ashenafi & Leader-Williams, 2005;Nigussie et al., 2019).Occupancy and detection probability estimates of Moorland Francolins were higher in traditionally protected areas than in unprotected areas, suggesting the persistent and high conservation effort supporting the Ethiopian Wolf (Canis simensis) by the local community in association with international organizations signifies the integrity and functionality of the whole community.Flagship species like this play a vital role in biodiversity conservation at local and global scales (Jarić et al., 2023), which is demonstrated by its positive side effects for Moorland Francolins and other species in GCCA, too.Unlike other carnivore species, this species is a rodent specialist (Ashenafi et al., 2012;Atickem & Stenseth, 2022;Vial et al., 2011).

| Determinants of occupancy and detection probabilities
Based on beta estimates and moderate model weight, Moorland Francolins revealed an aversion to montane forest habitat due to the presence of predators in the tree canopies.The Afroalpine highlands are suitable habitats for predators (Clouet et al., 2000), and habitat F I G U R E 6 (a, b) Occupancy probability (ψ) of Moorland Francolin in association with herb species richness and distance to the nearest road (km) and (c, d) Detection probability (ρ) of the species in association with sampling occasion and average precipitation (mm/day), respectively.The estimates for the parameters are created from the most parsimonious model that holds these covariates and the shaded area in each graph shows 95% confidence intervals.
use of many ground-dwelling birds is negatively influenced by the presence of predators in and around the forest habitats (Abrha et al., 2018;Sukumal et al., 2017).In concordance with these findings, our results confirm that predators (both aerial and ground predators) may strongly negatively influence the habitat use of Moorland Francolins in GCCA, although the main diet of several raptors is rodents (Clouet et al., 2000).
Though hunting pressure is one of the key factors for decreasing francolin populations nationwide (Abrha et al., 2017;Gedeon, Rödder, et al., 2017;Töpfer et al., 2014)  Herb species richness was also supported based on model weight and top models.The protected grassland of GCCA covers almost 60% of its total area (Steger et al., 2020) and holds several range-restricted species (Ashenafi et al., 2012;Ashenafi & Leader-Williams, 2005).As expected, the occupancy probability of Moorland Francolins increased with herb species richness in GCCA, in line with other reports on pheasant species (Jolli et al., 2012;Sukumal et al., 2017).This vegetation type is widespread in the plateau of Afroalpine biome of Ethiopia (Nigussie et al., 2019;Steger et al., 2020) and it is the source of food and provides essential shelter for many grassland specialists (Töpfer & Gedeon, 2020).It had also a positive influence on the habitat use of Moorland Francolins at SEA, but the 95% confidence interval of the β-coefficient estimate overlapped zero showing less support for its influence on the species.This is because the area has been increasingly transformed into a monocultural plantation (Bahru et al., 2021;Tadesse & Tafere, 2017), and is subject to tourism activities (Asefa, 2018;Tesema & Berhan, 2019), overgrazing and other human-induced disturbances in the plateau of central highlands (Asefa et al., 2020).For instance, a recent report showed that the natural grassland of Entoto Natural Park has decreased over the last three decades and that the area is now dominated by an Eucalyptus plantation (Tesema & Berhan, 2019).In such areas, Moorland Francolins showed a pronounced aversion toward modified habitat types.This implies that Afromontane grassland and shrubland specialists, especially Moorland Francolins might gradually become locally extinct.
Distance to road was also the other strongest covariate influencing the occupancy probability of Moorland Francolins, similar to other reports in ground-dwelling bird species (Whitworth et al., 2018).The occupancy probability of the species was higher along the edge of roadsides and trails than at sites located in remote in GCCA, in concordance with other reports on wildlife species (Kroeger et al., 2022;Paemelaere et al., 2023).This is unexpected because roads can attract hunters and predators, delivering also other human-induced perturbations (Dean et al., 2019;Kroeger et al., 2022).In GCCA, we observed that proximity to road attracts the species as there were food items mainly on the unpaved road, including grains and fruits thrown through window by passengers.Most roadsides have also dense native herbaceous vegetation, which may also help Moorland Francolins to survive.On the contrary, occupancy increased as the distance to road increased in SEA habitat but did not show a significant association with roads.This suggests that Moorland Francolins avoid roads and trails in a human-modified landscape.Thus, roads may have positive effects on bird species in more pristine habitats (Kroeger et al., 2022) and in areas where hunting pressure is controlled as a management strategy (Whitworth et al., 2018).
Local low temperatures and high ground vegetation cover (Nigussie et al., 2019;Steger et al., 2020) may lead the species to use the roadsides and trails: (1) to enhance foraging opportunities; (2) to stay more vigilant to avoid risk of predation; (3) as a heat source; (4) to facilitate mating, connectivity and communication.
Avoidance of human settlements is likely related to livestock grazing causing herb species richness to shrink at the GCCA periphery (i.e., human occupation).Similarly, the effect of distance to settlement as a type of human disturbance posed a positive effect on Moorland Francolins in SEA.There was no significant difference for the covariate in this area, yet relatively high model-averaged beta coefficient estimate; model weight and confidence intervals reveal irregularity in association with the species, most presumably due to lack of habitat heterogeneity, a small sample size, limited number of cameras, and small sampling occasions, as compared to recommended occasions.Hence distance to settlement had a slightly significant positive influence on the species in SEA, agreeing with previous studies on pheasants (Chen et al., 2019;Jolli et al., 2012;Nuttall et al., 2017;O'Brien & Kinnaird, 2008), other bird (Pardo et al., 2017) and mammal species (Paemelaere et al., 2023;Semper-Pascual et al., 2020).
In line with our hypothesis, sampling occasion significantly positively influences the detectability of the species in GCCA.
Conversely, in SEA, this covariate appeared in one of the most parsimonious models and it positively influenced detectability but it had low model weight and the beta coefficient estimates showed statistically non-significance association.The detectability may be affected by spatial variations and sample sizes.Our hypothesis that species detection increases with number of days of cameras deployed showed consistency with other findings in bird (Paemelaere et al., 2023;Si et al., 2014) and mammal species (Holzner et al., 2021;Semper-Pascual et al., 2020;Shannon et al., 2014;Si et al., 2014;Wevers et al., 2021).The magnitude of sampling occasion on detection probability estimate demonstrates species-specific response (Iannarilli et al., 2021).
In Ethiopia, after a long dry season, both a small and a main rain season occurs in most highland areas (Mohammed et al., 2022).Several francolin species are adapted to this seasonally changing precipitation regime (Abrha et al., 2018;Gedeon, Rödder, et al., 2017), which allows the areas to replenish food resources and ecosystem greenness vital for breeding (Abrha et al., 2018).This is because francolins may find plenty of food by easily raking and scratching the wet ground (Abrha et al., 2018).Moreover, during rain seasons, birds of prey soar less, and agro-pastoral encroachments seem lower compared to the dry season (pers.obs).Elsewhere in tropics, the breeding season of birds is reported to be associated with the beginning of precipitation and this is linked to the abundance in food and cover resources (Cox et al., 2013;França et al., 2020;Jansen & Crowe, 2005).In our species, some camera traps have documented chicks being fed by their parents in GCCA, and this implies that the breeding season of the species may coincide with the short and mild precipitation distribution from February to June.Similarly, temperature positively influenced the detectability of the species, but there was little support for our hypothesis based on models.This may suggest that the species avoids extreme temperatures.Collectively, climate factors are very important for the detectability of the target species in the central highlands of Ethiopia.

| Camera trapping for assessment of cryptic bird species
The Moorland Francolins, similar to other pheasants in the region, could potentially go visually undetected, particularly in areas of low population density and in disturbed habitats.Extreme weather conditions, seasonality, expert experience, and other factors may also obscure the ability of detecting the species.This is the birds usually remain silent, hidden, and squatted when people approach them.Thus, false-negative detection could bias inferences about the occupancy and detection probability estimates and other parameters.However, the deployment of non-invasive modern approaches like remotely triggered camera traps can avoid such ecological concerns.This approach also helps to discover new geographical ranges, other wildlife species (including predators) and thereby helping to understand the interactions of the Moorland Francolins in its natural habitat.Another positive feature of the camera trapping technique is that it is cost and time-effective.Our results strongly support the deployment of camera traps for the detection of cryptic and little-known species in a topographically complex region.Camera traps provide reliable comprehension and precision of occupancy of Moorland Francolins in the Afroalpine Biome.Such camera trap data (O'Brien & Kinnaird, 2008;Sharief et al., 2022;Si et al., 2014;Steenweg et al., 2017;Wearn & Glover-Kapfer, 2017) ultimately promotes the proper conservation of the target species.

| CON CLUS IONS
The findings demonstrate that habitat use of Moorland Francolins is higher in the more pristine habitats compared to the strongly humaninfluenced in SEA.This suggests that a community-based conservation area (i.e., GCCA) is a crucial remnant habitat of endangered and data-deficient wildlife species in Ethiopia.Since such communitybased conservation approaches obviously support sustainable species-habitat conservation, strengthening the existing Qero system and expanding the model to other potential hotspot sites and/ or IBAs is strongly recommended to circumvent the mounting anthropogenic disturbances in the region (Asefa et al., 2017;Chengere et al., 2022;Razgour et al., 2021;Rodrigues et al., 2021).
Our results also show that the species uses various herb species, roadsides and trails for resting, hiding, survival, and reproduction.Conversely, predators threatened the francolins predominantly in native and plantation forests, thus Moorland Francolins tend to avoid tree canopy cover and human settlements in both study areas.
In the human-modified SEA areas, most covariates had a weak influence on the occupancy and detection estimates of our target species because habitats are dominated by Eucalyptus plantations, fragmented meadow hill patches, and farmlands, unlike the heterogeneous and protected habitats in GCCA.
We confirm that camera trap deployment corroborates the presence or absence of shy ground-dwelling birds not only in known areas but also in understudied areas.The detectability of francolins was determined by the sampling occasion and precipitation.Further

ACK N OWLED G M ENTS
We are thankful to The German Academic Exchange Service (DAAD) for supporting the PhD project of AMA.We also thank Mekelle University, University of Bonn, and Leibniz Institute for the Analysis of Biodiversity Change (LIB) for financial and logistic support of our research.We are grateful to the intellectual guidance of Mr. Yilma Dellelegn Abebe for providing information on records of Moorland Francolins outside of the IBAs.We would like to appreciate the Ethiopian Wildlife Conservation Authority (EWCA) for granting research permission (Ref no: 31/74/12).We are very thankful for the help of GCCA staff and other volunteers during data collection.Last but not least, we wish to thank three anonymous reviewers for their helpful Afroalpine biome, camera trap, conservation, endemic, moorland francolin, occupancy T A X O N O M Y C L A S S I F I C A T I O N Agroecology, Autecology, Population ecology, Zoology F I G U R E 1 An adult female Moorland Francolin Scleroptila psilolaema in the Afroalpine biome, Ethiopia.The feather patterns contribute crypsis through background matching in this species (photo credit: Kai Gedeon).

.
Unlike other IBAs of the study areas, the Qero system, coupled with the conservation initiatives of Frankfurt Zoological Society, The Darwin Initiative, European Union, and Ethiopian Wolf Conservation Program have significantly sustained the ecological integrity of GCCA since 2003.In this area, the Ethiopian Wolf Canis simensis is the flagship species (Tefera & Sillero-Zubiri, 2006), generating income through ecotourism which is partly plowed back F I G U R E 2 The two study areas (GCCA and SEA) and location of camera sites in the central highlands of Ethiopia.GCCA, Guassa Community Conservation Area.The southern sites (including Sululta, Entoto National Park, Ankober-Debresina escarpment, and other areas) form SEA.
et al., 2018; MacKenzie & Royle, 2005), we initially attempted to conduct a total of 185 camera sites (or preferably sites) (n = 116 for GCCA and n = 69 for SEA) for a single-season design located in various habitat types.All camera sites were arranged in 39 line transects (n = 98, 2-10 days) to obtain an average of three occasions per site at GCCA area.Whereas cameras at SEA area were active for approximately eight consecutive days(n = 48, 4-12 days)  to obtain an average of four occasions per site.Number of camera days varied depending on the probability of detection of the species in the two different areas.Such study duration is recommended for high detectable species(Guillera-Arroita et al., 2010;MacKenzie & Royle, 2005).To account occupancy model assumptions(MacKenzie et al., 2002(MacKenzie et al., , 2018)), each site was surveyed between one to five repeated occasions (Κ max = 5; Κ average = 2.95) in GCCA from March to June 2020, while in SEA each site was surveyed two to six repeated occasions (Κ max = 6; Κ average = 3.46) from February to March 2020.The discrepancy in number of occasions per site was due to accessibility, logistical constraints, security, weather conditions, and technical problems.We had missed observations in some sites meaning that sampling was not conducted at site i during time t and hence a missed observation represented by hyphen (−) was filled instead in the complete detection history (h i ).
10,000 permutations to assess the adequacy of fit of the global model (i.e., the most parameterized model) and Pearson's Chisquare test (χ 2 ) and non-Bayesian p-value were implemented to check overdispersion (ĉ) (MacKenzie & Bailey, 2004).The degree of overdispersion parameter estimate (ĉ) or variance inflation factor was assessed using chi-squared (GOF) statistic.It was calculated by dividing the observed test statistic by the average of simulated test statistics.
ΔAIC c .The bootstrapping procedure and χ 2 test revealed that the global model (ψ(WD + Hsp + T caco + Pre + Elev + DR + DS + DW), ρ(E + T + P)) lacks overdispersion (χ 2 = 35.95;p = .35;ĉ = 0.85), showing independence among sites.Subsequently, the combinations of occupancy and detection covariates of the top models were tested based on the lowest ΔAIC c values.The bootstrapped top 13 models also showed adequate model fit (ĉ ~ 1, Table 2).The summed weight of the top-ranked models (ΔAIC c ≤ 2.0) was 0.61 and the most parsimonious model (ψ(Hsp + T caco + Pre + DR),ρ(E + T + P)) had only 0.08 model weight, suggesting more plausible competing models existed to explain the occupancy and detection estimates (

F
Parameter estimates (occupancy and detectability) of Moorland Francolin derived from model averaging.The asterisks (***) denote a strong statistically significant difference between parameter estimates in the study area at p < .001level.

F
I G U R E 5 Occupancy probability of Moorland Francolin in association with predator presence/absence in GCCA.Cameras placed in woody plant species frequently had photos of predators like Serval Leptailurus serval.Error bars indicate standard errors of occupancy probability, ***p < .001.
of overall occupancy(Guillera-Arroita et al., 2014;MacKenzie et al., 2018).Since we had small sample sizes and low density of individuals in SEA, we increased the sampling by one more occasion to minimize the effect of false-negative detections of the target species.Increasing of sampling occasion helps to increase the precision and accuracy of detectability of species (MacKenzie & Royle, 2005; Moore et al., 2014).

TA B L E 1
Habitat covariates predicted to affect occupancy and detection probabilities of Moorland Francolins in the central highlands of Ethiopia.

Covariate Type of data Measurement and scoring systems Hypothesized relationship References tested the effects
Results of model selection for Moorland Francolins occupancy and detection probabilities in the central highlands of Ethiopia.Model rankings are based on the AIC c values; AIC c values compared to the top-ranked model (ΔAIC c ); ΔAIC c scores ≤ 2.0 are the top-ranked model; model weight (ω i ), and number of parameters (Κ), and −2 L = −2Log e L. ĉ = overdispersion parameter to estimate lack of fit.
Moorland Francolins were detected at 68 of 98 sites, which resulted in a naïve occupancy (proportion of sites that recorded at least one photograph on the whole camera TA B L E 2

Table 2
Summed model weight (Σω i ) and influence of covariates calculated from model-averaged beta coefficient estimates and standard errors (β mean ± SE).
).We used model averaging to improve inference as the top model clearly Note: Lower and upper 95% confidence intervals of the coefficients were constructed.Nonoverlapping with zero (bold) shows significance values of β estimates.TA B L E 3