ANNEX 5: Pilot Project base maps
- Ituri/ Okapi Wildlife Reserve
- Odzala National Park
- Central Gabon / Lope Reserve
ANNEX 6: Site reporting form for Mike Intensive Survey Sites
MIKE SITE REPORT
TRIMESTRIAL REPORT FOREST
SECTION A: SITE RESOURCE DETAILS
A1. Name of site:
A2. Reporting period: first month: last month:
A3. Date form completed:
A4. Reporting officer:
A5. Expenditure for field patrols:
Amount in local currency
Salary and supplements
Field allowances / night out allowances
A6. Exchange rate and date: 1 $US =
A7. Patrol / surveillance transport used this period:
A8. Patrol equipment used/expended:
SECTION B: SEARCH AND MONITORING EFFORT
B1. Law Enforcement Monitoring Form
Were there any law enforcement activities during the last three months? YES / NO
Has a law enforcement monitoring form been filled out and attached to this report? YES / NO
B2. Patrol / Survey summary
No. of men
No. of man days
No. of km covered
No. of grid squares covered
Mobile (Mainly foot)
Mobile (Mainly boat/canoe)
Mobile (Mainly vehicle)
B3. Geographic coverage
Attach a map to show all patrol routes (including research and surveys) and location of barriers. Indicate frequency of grid square coverage.
Map attached? YES / NO
B4. Note any major changes in search effort and geographic coverage this trimester compared to previous periods and reasons:
B5. Monitoring effort
Have monitoring teams been active in the area? YES / NO
Have monitoring forms (recces/transects) been filled out and attached to this report? YES / NO
Is there a map that shows the activities of the monitoring teams this last trimester? YES / NO
If yes, is this map attached to this report? YES / NO
B6. Summary of transects and recces
SECTION C: ILLEGAL KILLING OF ELEPHANTS
C1. Killed elephants
A) Summary of all reports of incidents of illegal killing
Source of data
B) Summary of all incidents of contact with poachers and information on seized or recovered ivory, elephant meat and arms
Poachers reported through intelligence
Ivory seized or recovered
Arms seized or recovered
Elephant meat seized or recovered
C2. Number of elephant carcasses recorded:
Cause of death
C3. Attach a map showing locations of carcasses detected. Map attached? YES / NO
C4. Attach copies of the carcass forms filled out during this trimester.
Number of forms attached:
C5. Information concerning mortality of elephants due to natural factors (drought, disease, accidents, predators…):
C6. Factors linked to illegal killing:
Present but not important
Effects of conflicts and anarchie
Military presence and operations
Presence and operations of police
Activities of other authorities
Other economic activity:
Other economic activity:
Other economic activity:
Other economic activity:
SECTION D: HUMAN IMPACT AND ILLEGAL ACTIVITIES
D1. Is there a map that shows observations of indirect signs of illegal killing of elephants that are summarized in the table below? YES / NO
If yes is this map attached to this report? YES / NO
D2. Observed signs of poaching:
Recces and transects
Other sources (including intelligence)
No. of incidents
No. of groups / incidents
No. of groups / incidents
Suspicious vehicle tracks
Gunshots heard (large caliber)
Other (give details)
D3. Other notes:
SECTION E. Human elephant relations (Village Surveys)
E1. Were village surveys conducted during the period? YES / NO
IF yes, in how many villages where surveys conducted?
Is there a map locating surveyed villages? YES / NO
If yes, is the map attached to this report? YES / NO.
E2. SURVEY RESULTS:
Check (X) the appropriate cell for each survey location.
Record only activities or evidence reported for the site during the trimestrial period.
(km FROM SETTLEMENT)
Elephant crop damage
Evidence of elephant poaching
< 5 km
> 5 km
1 - 5
ANNEX 7: Spatial modelling of Odzala elephant dung encounter
Len Thomas, Rene Beyers, John Hart
Animal densities are generally related to environmental, topographical, habitat and other spatial variables. These spatial variables can be obtained from a combination of sources such as remote sensing, maps, GIS coverages and field data.
Spatial modelling enables the analysis of animal densities, in our case elephant dung densities, in relation to spatial variables. Explaining variation in animal densities with spatial variables is useful for management of wildlife and protected areas.
Spatial modelling also allows the estimation of animal densities of a sub-area of interest within the global study area.
Under certain conditions (good coverage of the area by survey units, selection of the appropriate variables) animal densities can be estimated using non-random surveys. In marine monitoring programmes for example, surveys are often conducted from ferries, merchant vessels, tourist cruise ships etc. because hiring a dedicated survey ship is very expensive. These "platforms of opportunity" follow non-random routes and data cannot therefore be analysed in a design-unbiased framework. In this case spatial modelling provides a valid alternative.
Analysis of Odzala data
For this example we used data from 44 recce-transects in Odzala (Figure 1). The survey area was divided into 3 non-continuous zones. We used total dung count at each sampling location as the response variable. The following potential covariates were obtained from a GIS:
- x - longitude of transect
- y - latitude of transect
- park - distance (in km) from park boundary. Positive values are outside the park, negative are inside
- vill - distance from nearest village (villages are thought to negatively impact elephants through poaching)
- cvill - distance from nearest "non-conservation" village. A conservation village is one where there is anti-poaching activity (e.g., an education programme), so this covariate identifies only those villages that could be sources of poaching
- guard - distance from nearest park guard camp
- patrol - distance from nearest regular park guard patrol route (much of the park is not patrolled regularly)
- road - distance from nearest road (roads are thought to negatively impact elephants through poaching)
Scatterplots of the count (z) against the covariates (Figure 2) showed that that some of the potential covariates were strongly correlated: vill and cvill; guard and patrol. There is little point in including both of these pairs, so we dropped vill and guard from the analysis.
In the first stage of the modelling several models were constructed using the covariates and the best model was selected. In the second stage dung counts were predicted over the area of interest.
We used a generalized additive model (GAM) approach. This type of modelling allows for smooth non-linear relationships between covariates and the dependent variable (dung counts). The modelling was done with S-plus software.
We started by fitting a full model that included all selected covariates. However some variables did not contribute significantly to the fit. To reduce the number of variables, an automated stepwise regression was used, using the Akaike Information Criterion (AIC) as the model selection criterion.
AIC is a quantitative method for model selection. It includes a term that measures how well the model fits the data and a second term that is a penalty for the addition of parameters. Model fits can be improved and bias can be reduced by adding parameters but at a cost of adding model complexity and increasing variance (Buckland et al. 2001). AIC provides a trade-off between variance and bias. The model with the lowest AIC is selected.
Several GAM models were generated in S-plus with different covariates and degrees of freedom. The model with the lowest AIC was retained (Figure 3) and contained the following covariates: distance from road (smooth non-linear relationship with 2.9 effective degrees of freedom) and distance from patrol routes (linear relatationship).A gam model is constructed, selected….
Expected dung counts at each location predicted
Predicted dung counts across the whole survey area (map)
Predicting dung counts at survey locations
We can use the above model to predict the expected dung count at each survey location, and use this in place of the observed value of counts in estimating total count (n) and variance in count (var(n)). One property of predictions from a generalized linear or additive model is that they have the same mean (and therefore total count) as the original data – so this procedure may seem of little use at first glance. However, if the model has explained some of the variation in counts, then var(n) will be smaller than the observed variance of the counts. Since var(n) is the major component of variation in the distance sampling estimates of density, this can provide a method of substantially reducing the variance of the density estimates.
The CV of raw counts is 14.6%, while the CV of the predicted counts is 9.6% - 65% smaller.
Predicting dung counts over the survey area
We can also use the model to predict dung count over the whole surveyed region. This is useful: (i) if we have no design-based estimate available (e.g., platforms of opportunity); (ii) for showing our model as a map; (iii) for estimating density over any sub-region of the study area.
To predict over the whole region, we need covariate values at a grid of squares (or other shapes) of known area over the region. In this analysis, we used a grid with cells of an area of 21.25km2 after projecting cells of 2°30’’ by 2°30" in Lambert equal-area azimuthal (equatorial). Covariate values were calculated at the center of each grid point.
The mean predicted count over the survey area will no longer be the same as that at the survey locations – it is 41.1 as opposed to 45.1. Since the survey was laid out using a stratified random design the estimate of 45.1 from the survey locations is unbiased; however the model-based estimate of 41.1 is only unbiased if the model is correct.
We cannot use a simple analytic formula to work out the variance and CV of predicted counts, because the estimates for each grid point are not independent. Instead, we usually use computer-intensive methods such as the jackknife (if there are many sample locations), or a parametric bootstrap. These are beyond the scope of this report.
A map of the predicted densities is shown in Figure 4.
Discussion, conclusions, recommendations
Interpretation of the results is outside the scope of this report. However, it is interesting to note that one of the selected covariates was distance from patrols: dung density appears to decrease with increasing distance from regular patrols. There was little dung seen at sample locations in the southern zone (Figure 3) – these are outside the park.
The largest residual value comes at the sampling location in the middle zone with an observed count of 210, and predicted count of 120. This sampling location was adjacent to the park headquarters, which may explain the high density of elephants. However, the closest other sampling location had an observed count of 34, it the high residual may represent real variation in density.
There is also a pattern of positive residuals in the middle 4 sampling locations running north-south in the northern sampling zone (Figure 5). Possibly some missing covariate (such as terrain or habitat) could be used to explain this pattern.
There is plenty of scope for future work. This includes:
- incorporating detection function modeling into the analysis, to allow for variation in delectability within the study area and allow prediction of density, rather than encounter rate.
- including travel recce data. The modeling would then follow either the count or waiting areas methods of Hedley (2001, unpublished PhD thesis).
- including further covariates, such as terrain and habitat.
- deriving variances for the model-based estimate using a computer-intensive method.
Buckland, S.T., Anderson, D.R., Burnham, K.P., Laake, J.L., Borchers, D.L. and L. Thomas. 2001. Introduction to Distance Sampling. Estimating abuncance of biological populations. Oxford University Press.
Hedley, S.L., Buckland, S.T. and Borchers, D.L. in prep. Spatial models for line transect sampling.
Figure 1. Map of sample locations, with count at each location.
Figure 2. Scatterplots of the covariates.
Figure 3. Fitted covariates, standard error of the fitted line, and residuals from the model selected by the automated stepwise regression procedure.
Figure 4. Covariates and density map of abundance estimates
Figure 5. Residuals from the model shown in Figure 3.
|Site map||Search the site||FAQ & contact us||Home|