- Complexity and Resilience
- Fresh Video
- Recommended for you
- lemurs ecology and adaptation developments in primatology progress and prospects Manual
- Spider Monkeys in Human-Modified Landscapes: The Importance of the Matrix
The objectives of our study were: 1 to obtain species-specific models from distance sampling data, using an approach that has rarely but successfully been applied to derive the abundance of endangered animal populations [ 20 , 28 , 29 ]; 2 to assess if such selected models could improve the sensitivity of estimates of primates population density; 3 to gain relevant information for conservation purposes by modelling the spatial variation of primate density in a highly heterogeneous and complex human-natural system.
Data collection did not involve direct contact or interaction with the animals. We analyzed data collected by earlier studies [ 26 , 27 ] in respect and under permissions of the relevant authorities indicated therein.
Complexity and Resilience
The mountains are characterized by the presence of several forest blocks that differ in elevation range —2, m a. Data were collected by Araldi et al. The study focused on three species of arboreal primates that show a widespread distribution across the Udzungwa Mountains: the Peters' angolan colobus Colobus angolensis palliatus henceforth BW , the endemic and endangered IUCN, Udzungwa red colobus hencefort RC and the Tanzania Sykes' monkey Cercopithecus mitis monoides henceforth SY.
We refer to Barelli et al. We used data in Araldi et al. Authors achieved a uniform coverage of target forests Fig 2. Arboreal vegetation and disturbance parameters were collected by establishing four squared vegetation plots, 25 by 25 m each, centered on each line transect, with a total of plots sampled see [ 27 ]. Map of Mwanihana forest MW with the sampling grid, as an example of diffused grid of transects walked for primate density estimations in Udzungwa Mountains National Park of Tanzania. We modeled the observed data as a hierarchical coupled logistic regression.
One step of the modelling process is related to the partially observed true state occurrence, the result of a biological process ; the other step describes detection, that is the result of both the biological process and the observation process i. We modelled the detection process according to a multinomial distribution and we expected the detection probability to monotonically decrease with the increasing distance from the observer, as per conventional distance sampling theory [ 24 ].
We verified this process by looking at the histograms of the distance records. We removed outliers from the data-set, defining a species-specific right- truncation distance, looking at the right tail of the plotted distance frequency distribution. Observations taken at larger distances were scarce and provided little information for the estimation of the detection function [ 24 ]. We therefore grouped in intervals distances that were recorded on a continuous scale, correcting for heaping and to improve estimates of density and better model fit [ 24 ]. Thus, we defined 5 bins of 20 m for the analysis on BW, 6 bins of 15 m for the analysis on RC and 4 bins of 25 m for the analysis on SY.
We first checked the performance of different detection functions uniform, half-normal and hazard-rate on the simplest model, without considering the covariates effect.
We built models using all the possible combinations of environmental and human-disturbance variables, sampled at the transect level, to determine how they affect both the detection process and the presence of the animals, based on an set of assumptions Table 1 ; see also [ 27 ]. In addition to distance which is an inherent covariate of the detection process, we assumed detection to be influenced by the following covariates: 1 group size, assuming that larger groups are more easily detected in the canopy at larger distances [ 24 ]; 2 forest block, as a nominal covariate representative of the heterogeneity among forests, given that each forest is a discrete area sampled; 3 canopy cover and 4 percentage of climbers; 5 steepness and 6 distance to anthropic disturbance i.
We used these covariates also when modelling the state process, in addition to 7 altitude, 8 diversity of tree communities, calculated as the Simpson's reciprocal diversity index; we also considered, as proxies of disturbance, 9 count of signals of human presence along transects cutting signs, recent and old paths, and trails made by humans, sites where pit sawing had been carried out or charcoal was produced, as well as signs of recent and old poacher camps, incidence of animal snares and 10 distance from the forest edge.
This criterion prevent us from unequivocally define a single best model on which to base predictions. We thus determined Akaike weights wi for each of the best models R package MuMIN ; [ 38 ] and to further reduce ambiguity, we derived the relative importance of each variable, on a scale from 0 to To verify the goodness of fit of the selected model we performed a parametric bootstrapping, simulating datasets from the fitted model and defining a function that returned the fit-statistic of the Pearson's X2. We used non parametric bootstrap to estimate the uncertainty i. SE of the parameters in the model.
Recommended for you
We then used the resulting best species-specific models selected, to predict primates group density, as well as their detectability, in each sampled forest block and in each of the plot that were sampled along the transects, for which measurements of the influential habitat variables were available. We also assessed how the hierarchical structure of our analysis could improve our estimates, by comparing our results with those from Araldi et al.
To test for differences between the two approaches, we used a t-test after assessing normality with Shapiro-Wilk tests [ 39 ]. After right truncating the data at m we retained 90 observations for BW and for SY, while we retained 97 observations for RC with a 90 m truncation. Detection functions indicated that all assumptions for the method were met, i. No spikes were present after binning the data in distance classes.
Detection functions from the best AIC models, shown for the 0. Covariates effect on group density estimation, shown for the best model selected for the Udzungwa red colobus RC. Testing for differences between density estimates from our null model and estimates in Araldi et al.
lemurs ecology and adaptation developments in primatology progress and prospects Manual
Spatially-explicit maps of estimated density at the plot level are shown in Fig 6 and Figure C in S1 File. Predicted values were obtained for the plots that were sampled along the transects, for which exact values of the influential covariates were available. Our study aimed to show the importance of accounting for habitat covariates of primate detectability and abundance in distance sampling studies in complex landscapes. The hierarchical analytical approach allowed us to obtain reliable, informative and spatially-explicit estimates relative to previous studies that did not consider the covariate effect [ 26 ] nor abundance estimation with imperfect detection [ 27 ].
Moreover, the method we used allows for inference on density outside the sampled area. This is of particular relevance when the variables retained in the modelling are spatially diffused, as it usually applies to those derived from remote sensing. A first important result is how the species-specific group size influences detection. By using this approach, group size effect could be explicitly evaluated and therefore modelled.
On the contrary, in conventional distance sampling group size is regressed on estimated probability of detection. The positive relationship between group size and detection in BW and SY, but not RC, is likely explained by different grouping patterns. The average group size of BW and SY was indeed similar 3.
Groups of RC could have been consequently more easily detected even far from the transect line. Thus group size represents a critical parameter that needs to be carefully considered to avoid underestimation of animal densities, with particular relevance for species whose social units are small i. As predicted, we found detectability for RC to be negatively influenced by distance from disturbance.
This variable represents a proxy for forest structures that can hamper visibility, such as tall and dense canopy in interior forest. Climber percentage, on the contrary, had a positive association with RC detectability.
As for the effect of covariates on animal density, we found the percentage of climbers to have a positive effect for all the three species we examined. This result is in line with findings from Barelli et al. We found a negative association between altitude and density of RC and SY. This also matches the findings from previous studies [ 27 , 41 , 46 ] that explained this result in terms of different food availability along the elevation gradients of the study area.
Human impact was found to have a negative association only with BW. Hunting pressure is indeed reported to be targeted mainly on this species, which skin is highly demanded [ 11 ]. RC and SY appear less affected by hunting and this differential degree of human impact is reported in several other studies [ 11 , 47 — 49 ]. Density of RC was related to the mean basal area of trees, that had a positive effect, and to distance from disturbance, with a negative effect, contrary to what we hypothesized. This is in line with results by Rovero and Struhsaker [ 41 ] and confirms the preferences shown by the species for larger trees that can be found also at forest edges.
Here, even if logging is more intense, productivity of the remaining large trees can still be high [ 50 ], thanks to an increase in illumination [ 51 ].
Nevertheless, variation in density between US and the other forest blocks was particularly substantial for BW Table 4 , for which the parameter level US was found to have a high negative effect on density estimation. Importantly, the variation in density estimates among forests was almost two times lower than that reported in Araldi et al. Such underestimation may have been smoothed by our analysis because of adding the effect of covariates on both the detection probability and the state process [ 52 , 53 ].
This is of particular conservation relevance in highly disturbed habitats, like US, where animals are sparse and shy, and therefore tend to hide and go undetected relatively more than in other forests Table 4.
In general, our results further confirm that the absence of protection in US highly impacts the colobine monkeys, with pressures that mainly derive from targeted hunting and to lesser extent to habitat degradation [ 11 , 54 ]. These findings in turn support the hypothesis that colobines are more sensitive than Tanzania Sykes' monkeys to highly disturbed habitats and to human impact that deeply affects the structural characteristics of the forest [ 54 — 56 ].
Obtaining reliable and informative estimates of primate density in complex and human-modified landscapes is difficult, yet with habitat degradation and loss being a pan-tropical phenomenon, an increasing proportion of primate species is found in degraded and patchy habitats [ 15 ]. Our study demonstrates how the inference on abundance is improved by accounting for habitat covariates as separately affecting the observation and the state processes. Indeed when compared to the canonical approach to distance sampling, the method we used refined density estimation differences among forests.
We also showed that group size influences the observation process and is of particular importance for species or populations with small social units. Lastly, this approach allows spatially-explicit modelling of animal density at the scale of the covariates used in the modelling.
Hence, when significant covariates are available across the study area forest blocks in our case , and even beyond, such as from remote sensing layers e. We thank two anonymous reviewers for constructive comments on the manuscript.
Spider Monkeys in Human-Modified Landscapes: The Importance of the Matrix
We thank C. Garzon, V. Andreo and S. Tenan for valuable comments on earlier versions of the manuscript. Conceived and designed the experiments: NC FR.