French Guianan mammal and bird population densities with spatial-capture recapture, line transect distance sampling, and 'unmarked' density models
Data files
Jan 08, 2025 version files 27.41 MB
-
activity-data-previous-surveys-2025.csv
227.25 KB
-
calibrated-cameras.rData
3.56 MB
-
calModel_8A_2024.rData
139.94 KB
-
dataTable_2025.csv
23.48 MB
-
README.md
3.22 KB
Abstract
Estimating population densities of large wildlife in forests has traditionally relied on either spatial capture-recapture (SCR) for animals that are individually marked, or line transect distance sampling (LTDS) for species observable by researchers. Recent advances have now introduced more general camera trap models that are applicable to a large range of terrestrial medium to large-sized species. However, validation of these unmarked density (UD) models remains scarce due to the lack of available density estimates that can serve as references. We tested the accuracy and precision of three UD models (the random encounter model, camera trap distance sampling, and the time-to-event model) against independent SCR estimates for ocelots and jaguars, and against LTDS estimates for eight unmarked species varying in abundance and ecology. We found that UD model estimates for ocelots were relatively accurate in matching the estimates from SCR, albeit with less precision. Additionally, UD model estimates were similar to LTDS estimates for seven out of eight studied species. Estimates for jaguar were, however, not similar between UD models and SCR, implying that UD models are not useful for rare species. Overall, our findings show that UD models are promising tools for monitoring abundant to relatively rare unmarked species in forests, although SCR remains preferred for marked species. To ultimately aid conservationists in making informed conservation decisions, efforts should be directed towards enhancing the precision of these models as well as their ease of access for non-technical practitioners.
README: Validating density estimation methods for unmarked wildlife with camera traps
https://doi.org/10.5061/dryad.fttdz091v
Description of the data and file structure
The scripts for the unmarked density models are provided as follows:
- o image-processing-unmarked-2024.R: this script takes as input the annotated image database (which we made in the TimeLapse2 sofware) as a.csv and processes the images. The camera model is calibrated following the CTtracking protocol and various unmarked parameters are estimated such as encounter rate, activity level, and animal movement speed.
o unmarked-models-2024: this script takes the output from the former script and uses the parameters to estimate the effective detection zones and eventually the densities with CTDS, the TTE model, and the REM.
The data that are required as input for the first script are provided; the input for the second script is the output from the first script. For the first script, we provide:
- The "dataTable_2025.csv" file with all the annotated data. 'NA' is used for values that are not applicable either to animal pictures or to non-animal pictures. The columns are as follows:
- Id, image_id, deploymentID: two identifiers of the image and one for the camera trap deployment.
- Empty: whether the photo does not contain a person/animal.
- Behavior: for animals, the depicted behavior (no animal = NA).
- bottom* *stick and top stick: in meters, the bottom and top measurements on the calibration stick that were used for calibration.
- start_date: date of installation of the camera
- end_date: date of recovery of the camera
- adj* *end date: final date of the data used (some later data was excluded to remain in a single season)
- effort: total number of observation days
- number of objects coords, bottom stick XY t, top stick XY t: these are the coordinate measurements as outputted by the TimeLapse2 software. number of objects coords indicate animal coordinates, bottom and top stick coordinates indicate the two measurements of the calibration stick.
- Calibration: whether the photo is used for calibration or not.
- first_detection: whether the photo concerns a first instance of an animal in the viewshed or not.
- group_size: the number of animals associated with this picture.
- ImageWidth / ImageHeight: photo size, in pixels
- The calibrated-cameras.rData file with the calibrated cameras following Marcus Rowcliffe's CTtracking protocol
- The CTtracking.R file developed by Marcus Rowcliffe, from his github. At some point in time this'll become an actual package.
- calModel_8A_2024.rData, which is an R object containing the calibrated camera model. Not required if using the calibrated-cameras.rData file
- activity-data-previous-surveys-2025.csv - additional encounter data from the survey site to help estimate activity levels. The rows are species detections with their timestamps.
Sharing/Access information
The CTtracking protocol and associated scripts can be found on Marcus Rowcliffe's github: https://github.com/MarcusRowcliffe/CTtracking
Methods
In 2022-2023, we conducted three surveys at a single tropical forest site in French Guiana: 1) a camera trap survey for unmarked density (UD) models for all terrestrial mammals and birds; 2) a camera trap spatial capture-recapture (SCR) survey for ocelot and jaguar; and 3) a line transect distance sampling (LTDS) survey for all large-bodied mammals and birds that can be observed by researchers walking on transects. The study site is a continuous block of coastal plain swamp forest contained within the boundaries of the Centre Spatial Guyanais (CSG) in French Guiana (RichardHansen et al., 2006; Petit et al., 2018).
Camera trap survey for unmarked density modelling
For the unmarked density models, we installed CTs at 150 deployment locations by rotating 50 camera traps (model BTC-8A) three times after a sampling period of approximately one month. CTs were installed on a systematic grid with 0.5 km interspacing at predetermined points to avoid selecting locations with biased encounter rates. We aimed to install cameras within 10 m of predesignated points and randomly pointed the device either north or south to avoid sun glare. Points were re-designated within 30 m if no suitable tree was present (e.g. due to water or a tree fall gap). Cameras were installed on tree trunks at a height of 0.3 m and configured to make 8 pictures per trigger with 1 s delay (the fastest option) at 8 MP resolution. The UD models require that the angle and radius to animals in the viewshed are known, so we calibrated our cameramodel following the camera trap tracking protocol (https://github.com/MarcusRowcliffe/CTtracking) and made approximately 20 pictures with a marked pole at various viewshed positions at each deployment location.
Camera trap survey for SCR modelling
For both ocelot and jaguar SCR, we deployed a second camera trap grid in 2022-2023 to estimate jaguar density through SCR. This survey essentially replicated the design of Petit et al. (2018), who previously estimated jaguar density in the study area by installing 70 paired CTs across the study area to observe both flanks of jaguars. Unlike the UD survey, devices were installed on trails to maximize detections, and with approximately 3 km interspacing between devices. Across both the UD survey and the SCR survey, which were deployed at the same time and location, we recaptured and identified enough individual ocelots to estimate densities for this species with SCR as well. The last camera trap survey used in this study was conducted by AUTHOR OF THIS PAPER, ANONYMIZED FOR PEER REVIEW in 2014-2015 to estimate tapir density through SCR, but it was inconclusive due to too few detections. This survey deployed 66 paired CTs with approximately 750 m interspacing. While inconclusive for its original purpose, it did yield many detections of other species, which we used as supplementary data for the activity level parameter required for the UD models (Supplementary Information).
Line transect distance sampling survey
We conducted an LTDS survey at the study site in December 2022 to compare density estimates from UD models to those derived from LTDS surveys. Trained observers walked a total of 179 km at a pace of ~1 km/h on 4 pre-cut trails, between 0700 and 1100 and 1430 and 1800, following standard line transect protocols (Denis et al., 2017). Perpendicular distances between observed animals and the transect were recorded to the nearest meter with a laser range finder. Observers noted group sizes for group-living animals as well as whether they were likely to have observed all members of the group or not; only ‘confident’ group size observations were used to calculate mean group size for these species.
Data processing
We identified ocelots and jaguars based on pelage patterns and constructed capture histories for individual animals. We fit these to a spatially explicit capture-recapture model with the SCR package for R 4.3.1 (Efford and Efford, 2023). Spatial capture-recapture (SCR) models estimate population densities by calculating the locations of animal activity centers (i.e. the midpoint of their home range) in a landscape (Borchers and Efford, 2008). The input for SCR models are recaptures of individual animals at known locations. It includes a state model describing the locations of the animal activity centers as well as an observation model describing the detection probability of animals at sensors (e.g. camera traps) as a function of the distance to their activity center. We fit hybrid mixture models with sex-specific capture probabilities (𝑔0) and ranging parameters (𝜎) and compared AICs to select the best model. The probability of detecting animals as a function of the distance to their home range center was modeled with a half-normal detection function.
With the line transect survey data, we estimated densities with the Distance package for R 4.3.1 (Miller et al., 2019) for all terrestrial taxa that were observedmore than once. These were: red-rumped agouti (Dasyprocta leporina), red acouchi (Myoprocta acouchy), South American coati (Nasua nasua), collared peccary (Pecari tajacu), black curassow (Crax alector), grey-winged trumpeter (Psophia crepitans), brocket deer (Mazama americana and Passalites nemorivagus) and tinamous (Tinamus major and Crypturellus variegatus/cinereus). Brocket deer and tinamous were each grouped when comparing UD models to LTDS models as these taxa were often difficult to identify to the species level during the LTDS survey. For the detection function, we supplemented the perpendicular distances observed on the transects with data from other transects in French Guiana (Denis et al., 2017). We then righttruncated the outer 10% of the detections and fit both half-normal and uniform distributions with and without adjustments, using the AIC criterion to select the best model. Encounters of multiple animals were modeled as if for a single individual only, and afterward we multiplied this ‘group density’ with the estimated mean group size, based on confident counts only.
Estimation of the snapshot interval
CTDS is based on the assumption that animals in front of the cameras are continuously captured on photographs, which requires estimation of the ‘snapshot rate’ of a continuously triggering camera. The average time in between photographs is the snapshot interval 𝑡 from (Howe et al., 2017). While 𝑡 can simply be calculated from the camera specifications given the manufacturer, these have often been found to be inaccurate (Corlatti et al., 2020). Most studies have therefore estimated an ‘empirical 𝑡’ by continually triggering a camera. Kühl et al. (2023) however found that in reality, such empirical 𝑡 values are still biased because 1) animals may not trigger cameras similarly to humans, and 2) animals are not continuously moving in the viewshed and therefore do not trigger the motion sensor instantly. In this study, we therefore estimate species-specific values for 𝑡 as simply the mean time interval in seconds between consecutive camera triggers of individual animals. Because individual animals cannot be identified, it is however difficult to determine whether individuals have actually remained in the viewshed in between triggers. A threshold 𝑇 is therefore determined, giving the time between triggers beyond which it is deemed more likely that the animal has left the viewshed before the second trigger than that it has remained in sight without triggering the device. A formal approach for estimating 𝑇 is not available yet (Kühl et al., 2023), and we here chose 𝑇 =20 by visually inspecting the photo data and making a subjective judgement on whether animals were more likely to have remained in place (e.g. animals sitting in the same place in both sequences) or to have left the viewshed in between sequences (e.g. animals moving near or towards the edge of the zone).
Estimation of the effective detection zone
The angles and radii between animals in the camera viewshed and the camera itself were derived following the CT tracking protocol (https://github.com/MarcusRowcliffe/CTtracking). We calculated 𝑟 and 𝜃 for all species by fitting hazard, half-normal, and uniform detection functions with and without polynomial, cosine, and hermite adjustments to the angles and radii of the positions at which animals were first detected by the camera. We only used the first photo of each detection for these functions, since the last seven pictures are always taken independently of the camera sensor response. We selected the best performing detection models with the AIC criterion.
Estimation of day range
To estimate day range 𝑣, we estimated species-specific movement speeds for sequences of photographs separated by less than 20 seconds as the cumulative displacement between animal positions divided by the time passed between pictures, and we averaged speeds to produce one value per independent detection. We discarded speeds above 1.5 m/s, which were deemed unrealistically high, since running animals were virtually never observed. To account for the fact that cameras are more likely to encounter fast-moving animals, we modeled speed with log-normal, gamma, or Weibull distributions following Rowcliffe et al. (2016), selecting the best performing model for each species with the AICc criterion. Because animals are not active at all times of the day, the movement speed observed by the CTs needs to be corrected for the time that the animals are actually active and are moving at those speeds. We therefore estimated species-specific activity levels (i.e. the proportion of time an animal populations is available for detection) by fitting a circular kernel model to the radian time data with the activity package for R (Rowcliffe et al., 2014; Rowcliffe, 2016). To ensure independence between observations, we only used observations separated by at least one hour for activity level. This approach assumes that all animals in the population are active at the time of peak activity for the species. Previous research has recommended a minimum of 100 independent detections to estimate activity levels (Lashley et al., 2018), which was not reached for curassow, ocelot and jaguar. We therefore supplemented the time-ofday detection records for all species with encounters and timestamps from all CT surveys combined. Including these data, we achieved 199, 232 and 96 independent encounters for curassow, jaguar and ocelot respectively.