The field survey was carried out from October 2011 to mid-June 2012 by trained researchers. Since extensive area was to be surveyed in the given time, three researchers were appointed to carry out the survey.
Prior to starting the survey, the study area was gridded as per the given grid size. Within each grid, potential sites to be surveyed were identified based on visible forest cover using Goggle Earth. The Forest Divisions within each grid were identified and after discussing with the local officers, the research team visited each site and verified the presence of tall forests within the division. The survey was carried out at the range level within each forest division.
After identifying potential sites to be surveyed in each range, three survey routes were selected per day. Each route was surveyed by one field researcher accompanied by a local guide. The survey for giant squirrel was carried out along existing trails in the sub-grids. The sub-grids were accessed by driving to the nearest point by a four-wheel vehicle and then walking to the start of the trail. The surveys were carried out through the day between 08.00 hrs to 17.00 hrs as per the availability of forest cover in the sub-grid. An attempt was made to start the survey early in the morning since this is the period when squirrels are most active. However this was not always possible since the sub-grids were often located far from the campsite. The length of trail in each sub-grid was about 2 to 3 km where there was substantial forest area. Correspondingly lower effort was employed in sub-grids where the forest area was lower.
We recorded giant squirrel presence based on direct evidences (sightings and calls), and indirect evidences (nests). Since calls are heard from a maximum distance of 100 m the squirrels were assumed to be in the same sub-grid as the observer even if recorded at the boundary of the sub-grid. Distance sampling was used for population estimation. Squirrels are generally found singly but are occasionally seen in pairs. On sighting the squirrel, number of individuals seen, perpendicular distance of the squirrels from the trail, and sighting angle were recorded.
The nests were categorized as new and old based on their appearance. New nests were lined by green foliage and appeared compact while old nests generally had a disheveled appearance. The nest trees and nest heights were also recorded.
Observations on types of threats encountered were recorded on the trail. All observations, including the direct sighting of squirrels, locations of nests and threat types were marked on the GPS.
In-use or new nest and un-used or old nest
Prachi Mehta Research team
List of Researchers who participated in the Survey
4.1 Determinants of Giant Squirrel Distribution in the Study Area Our objective was to examine potential covariates that influenced grid-level giant squirrel occupancy in the landscape; however we expected that some of these covariates would also influence squirrel detection due to variation in squirrel behavior, abundance or observer visibility. Therefore, we first prepared a set of candidate models to determine the covariate structure for detection probability p (replicate-level) and then further used this structure along with potential covariates to model squirrel occupancy Psi (based on Karanth et al. 2011).
Habitat Factors Considered as Covariates A total of 8 covariates were considered for the analysis and specific hypotheses based on their expected direction of influence were established. For each grid, we used a few remotely sensed variables and few field based covariates. The table below describes a list of potential variables, and expected direction of influence in explaining giant squirrel occupancy and detection probability in the study area.
We performed pair-wise Pearson’s correlations for the variables to examine if any variables were highly correlated with each other (r > ±0.6). Edge density per grid was highly correlated with Percentage forest cover and Edge density for patch, therefore it was removed from further analysis (Table 4.1) All covariates, except for the categorical variable for Presence of Protected area, were converted to z-scores and used in the occupancy modeling analysis further.
Table 4.1: Habitat Covariates used in Estimating Site-occupancy by R. indica
Area of forest cover for both classes of forest, semi-evergreen and moist deciduous, were calculated from the LULC MAP. This area was divided by the total area of the grid and converted to percentage.
Slope was calculated in degrees using 30m ASTER DEM raster for the study area. This was then averaged for each grid, using the ‘Zonal statistics’ plug-in in QGIS (Version 1.7.0 Wroclaw).
Presence of Protected area
Presence of Protected area was a categorical variable represented by a ‘1’ or ‘0’.
Mean Patch size
Patch size was calculated for patches of Forest (both semi-evergreen and moist deciduous) in FRAGSTATS at the level of every grid. Mean Patch size was calculated by dividing the sum of all Patch sizes by the number of patches within a grid.
Disturbance index was calculated as an encounter rate, which included the number of disturbance signs detected in each grid divided by the walk effort.
Edge density per grid
Edge density per grid was calculated by dividing total perimeter of Forest patches with total area of Forest patches for every grid. Here the Forest patch was restricted to within each grid cell, i.e. to say that area or perimeter of Forest patches extending beyond the grid cell were not calculated. Only patch characteristics within the grid cell were used for analysis.
Edge density for Patch
Edge density for patch was calculated by dividing total perimeter of Forest patches with total area of Forest patches for every grid. Here the Forest patch was NOT restricted to within each grid cell, i.e. area and perimeter of Forest patches extending beyond the grid cell were also calculated. Patch characteristics beyond the grid cell were used for analysis.
Walk effort was calculated in kilometers using the sum line lengths tool in the ‘fTools’ plug-in in QGIS (Version 1.7.0 Wroclaw)
Table 4.2: Pair-wise Pearson’s correlation explaining occupancy and detection probability for R. indica
% Forest cover = Percentage forest cover; Average Slope = Average Slope per grid; PA = presence of Protected Area; Mean Patch Size = Mean patch size per grid
We then ran another 10 alternative models, keeping the covariate structure for Psi fixed, to select the model structure for p. The lowest scored AIC model (highest ranked) was selected to fix the model structure for p. We then kept this model structure for p unchanged and further ran a total of 11 occupancy models. The covariate ‘Presence of Protected area (PA)’ was removed from the all the occupancy models since the parameter was facing convergence issues and causing lack of model fit during initial model runs. We re-ran all 11 occupancy models without the PA covariate and used an information-theoretic model selection approach (Burnham & Anderson, 2002).
Model fit was examined by bootstrapping and observing c-hat values. Model averaging was performed and weighted parameter estimates and unconditional standard errors were calculated for model parameter estimates from the best ranked models (AIC<2) (Burnham & Anderson, 2002).
All analysis was performed in software PRESENCE (version 3.0 beta, Hines 2006). Model selection was performed using Akaike’s Information criterion (AIC) and model weights. We first defined a global model for occupancy, which included all covariates except walk effort, which was only used to model its effect on detection probability.
The top ranked model (Akaike weight (wi) = 0.445) for the covariate structure for replicate level detection probability (p) has been shown in Table 4.3. Percentage Forest cover, Walk effort, Slope and Disturbance index were ranked as important variables in explaining replicate level detection probability. This model structure was chosen for detection probability (p) in further occupancy analysis.
As per this model, for squirrel occupancy Psi (% Forest Cover + Slope+ Protected Area + Mean Patch Size +Edge Density for Patch+ Disturbance index) variables were considered and for detection probability p (Walk effort + % Forest Cover + Slope + Protected Area + Mean Patch Size +Edge Density for Patch) variables were considered.
Table 4.3: Candidate models considered and top model (highlighted) to define the covariate structure for detection probability (p)
Where, FC = Percentage forest cover; Slope = Slope; PA = presence of Protected area; MPS = Mean patch size; ED = Edge density for patch; DI = Disturbance index; Effort = Walk effort; psi = site-level Occupancy parameter; p = replicate level detection probability parameter; (.) = null model; AIC = Akaike’s information criterion score; ΔAIC = delta AIC or difference in AIC score; AIC wgt = Akaike weights
After keeping this covariate structure fixed for p, we used 11 plausible models for examining their effect on site-level occupancy of giant squirrels. No single model received most support and three models were ranked as the top models (with delta AIC <2) (Table 4.4). Since no single model received adequate support, model averaged parameter estimates had to be calculated for these three top models. Model averaged parameter estimates and their unconditional standard errors (SE) have been shown in Table 4.5. All models had c-hat values close to 1, indicating good fit for models overall.
Table 4.4: Candidate models considered and model selection results (highlighted models) for explaining R.indica occupancy (Psi)
No. of Para-meters
Where, FC = Percentage forest cover; Slope = Slope; MPS = Mean patch size; ED = Edge density for patch; DI = Disturbance index; Effort = Walk effort; psi = site-level Occupancy parameter; p = replicate level detection probability parameter; (.) = null model; AIC = Akaike’s information criterion score; ΔAIC = delta AIC or difference in AIC score; AIC wgt = Akaike weights
Model averaged parameter estimates, along with 95% Confidence Intervals (CIs) indicated that Slope and Disturbance index both had a negative influence on occupancy of Giant squirrels as expected (Table 4.5). Other covariates in the top models, including Forest cover, Edge Density for Patch and Mean Patch size did not show up as important in explaining occupancy due to large unconditional standard errors for the parameter estimates. Walk Effort, % Forest cover, and Disturbance index were important covariates in explaining detection probability of giant squirrels.
Prachi Mehta Table 4.5: model averaged weighted parameter estimates and unconditional standard errors in explaining R. indica occupancy and detection probability
Wt. Parameter estimates
Wt. unconditional SE
Upper 95% CI
FC = Percentage forest cover; Slope = Slope; DI = Disturbance index; MPS = Mean patch size; ED = Edge density for patch; Effort = Walk effort; psi = site-level Occupancy parameter; p = replicate level detection probability parameter; Wt. Parameter estimates = weighted parameter estimate; Wt. unconditional SE = weighted unconditional Standard Error; Lower 95% CI = Lower 95% Confidence Interval; Upper 95% CI = Upper 95% Confidence Interval
4.2 Site-Occupancy of Giant Squirrel in the Study Area The naïve occupancy (Psi) that is generated without using the capture-recapture framework, was found to be 0.75, i.e. to say that 75% of the sampled landscape was detected to have squirrels. Final parameter of occupancy (Psi) was estimated to be 0.95 (SE = 0.03). The probability of detecting (p) squirrel presence, if present at a replicate was estimated to be 0.61 (±0.05).
Table 4.6: Model parameter estimates and standard errors from the best models in explaining Giant squirrel occupancy and detection probability
Where, FC = Percentage forest cover; Slope = Slope; DI = Disturbance index; PA = presence of Protected area; MPS = Mean patch size; ED = Edge density per grid; Effort = Walk effort
Figure 4.1: Giant Squirrel Distribution based on Predicted Occupancy Psi
Figure 4.2: Conditional Occupancy of Giant Squirrel in the Study Area
Figure 4.3: Grids showing Detected /Undetected History
Figure 4.1 shows predicted occupancy model of giant squirrel presence in Maharashtra Western Ghats. Here the occupancy is predicted based on the attributes of covariates for each sampled grid. Based on the covariate values of each grid and capture histories, the model predicts occupancy estimates as shown. This map predicts giant squirrel distribution in the area.
The conditional occupancy model of giant squirrel presence in the study area is as shown in Figure 4.2 In this map, the grid where the giant squirrel was detected was assigned a score of 1 (deep purple color). In grids where the giant squirrel was not detected the occupancy is considered as the Psi value estimated by the occupancy model. The map uses predictive values for un-detected grids. The map helps understanding the status (and therefore the cause) of associated covariates in the undetected grids.
Figure 4.3 is prepared plainly with detected and undetected (1 or 0) records based on the field data. This map displays distribution pattern of giant squirrel at the landscape level.