Resource Selection by African Buffalo (Syncerus caffer) in the Caprivi Strip, Namibia
The Caprivi Strip is located in Northwest Namibia (figure 1). It is a narrow region,
only just over 30km across in some areas. The Caprivi is bounded to the North
by Angola and Zambia, to the East by Zimbabwe, and to the South by Botswana.
There are many major rivers running through the region, including the Zambezi,
Okavango, and Chobe. The Caprivi is quite flat, and is characterized mainly by
grassland, woodland, and savannah habitats (Naidoo et al. 2011 in review).
The climate is tropical and there is large variation in rainfall; the annual
average of 650 mm per year falls mostly between November and April
(Naidoo et al .2011 in review).
1. Data collection Figure 1. The Caprivi Strip is located in the Northeast of Namibia (Inset).
GPS fixes were downloaded for 32 buffalo at 5-hour intervals, from October 2007 - November 2011. These fixes included X/Y location, Longitude/Latitude and Date/Time.
Satellite data for Evaluated Vegetation Indices (EVI) in the Caprivi Strip were downloaded from NASA REVERB ECHO . The specific data downloaded was: MODIS/Terra Vegetation Indices 16-Day L3 Global 250m SIN Grid V006, for 2007-01-01 - 2011-11-02.
All GIS layers were obtained courtesy of the Ministry of Environment and Tourism, Namibia (table 1).
Table 1. GIS layers used in the analysis. All layers are Figure 2. The GIS layers used in the analysis.
for the Caprivi Strip, Namibia.
2. Data processing
The EVI data had to be processed before analysis. To do this, I used the 'Extract by Subdataset' tool (Spatial Analyst toolbox) to extract all the subdataset 1's for each EVI file. I reprojected the data into WGS 1984 UTM 34S. I also used the 'Extract by Mask' tool (Spatial Analyst toolbox) to clip the EVI images to the study area, to cut down processing time.
For all other GIS layers, I reprojected all data in WGS 1984 UTM 34S.
1. Generating home ranges
I used a script in R to generate Local Convex Hull polygons representing a home range for each of the 32 buffalo (Using the GPS fixes). These polygons were imported to ArcMap.
I added the GPS fix data to ArcMap and dispayed the XY data, along with the home ranges (figure 3).
2. Generating a random data points
In GME, I generated shapefiles containing randomly located points for within each home range, so that each set of randomly located points contained the same n as the GPS fixes for that home range. I then generated XY location data for each random point, also in GME. I added the resulting shapefiles to ArcMap and used the 'Merge' tool (Analysis toolbox) to generate one single layer for random points (figure 4).
3. Building the data sets
For both the GPS fixes and random points layers, I used the 'Near' tool (Analysis toolbox) to generate distance to the nearest road, fence, river, channel, pan, field, and barrier (a combination of roads, fences, and rivers). I used the 'Spatial Join' tool to select the vegetation type (from the Vegetation Structure layer) for each of the GPS fixes and random points. Finally, I used the 'Extract by Points' tool (Spatial Analyst toolbox) to extract the fraction tree cover for each of the GPS fixes and random points. I then extracted the attribute tables for both the GPS fixes and random points layers to CSV files.
In R, I ran a script to extract the EVI value for each GPS fix and random point, by location and date. For the random points, I calculated the EVI by generating a random date for each point using the date range from the associated home range ID. For both datasets I also calculated whether each data point was recorded during day (from sunrise-sunset) or night. I then exported these new variables to CSV files and combined them with the CSV files exported from ArcMap using Microsoft Excel.
In Excel, I created a CSV file for each home range ID that contained both the GPS fixes and random points for that home range, and all other covariates. For the vegetation structure, I generated dummy variables as the data were not numeric. I added a 'Presence' column, and for the GPS fixes, assigned a value of '1' and the random points a value of '0'.
4. Logistic regression
In ArcMap, I examined the location of each home range in relation to the variables I was interested in testing in order to decide which variables to test for each home range. Not all were applicable, e.g. I only had pan data for the West Caprivi, and only channels in the East Caprivi. I recorded these qualitative observations in a table for reference when conducting logistic regressions in R.
In R, I ran scripts to conduct logistic regressions for select independent variables vs. 'Presence' (dependent variable). I loaded each home range-associated CSV in separately, calculated correlation matrices, removed correlated variables, and then used the glm function to run a logistic regression on the variables of interest for each homerange, from my qualitative observations. I saved the results of each regression.
5. Displaying the results
To display my results spatially, I chose one buffalo ID whose homerange included interesting features, and generated probability surfaces for two different dates, one in the wet season, and one in the dry season. To do this, I converted all my layers to raster, and recalculated distances, as outlined above, using the 'Euclidean Distance' tool. I used a resolution of 250m for all my rasters, as my EVI data were in 250m resolution. I then used raster calculator to evaluate the logistic regression equation. I used raster calculator on this new raster, to perform a logit transform, in order to generate probabilities between 0 and 1.