Introduction
Zoometric data derived from unmanned aerial vehicle (UAV)-based wildlife surveys are attracting increasing attention because of the ability to collect data on the condition of individual animals. Surveying species to obtain accurate population estimates is a necessary but challenging task requiring a considerable investment of time and resources. Traditional ground-based monitoring techniques, such as camera traps and surveys performed on foot, are resource-intensive, potentially inaccurate and imprecise, as well as being challenging to validate (Gonzalez et al. Reference Gonzalez, Montes, Puig, Johnson, Mengersen and Gaston2016). For conservation purposes, it is essential to collect consistent and reliable information about species distribution and abundance to develop plans for species protection and sustainable population management (Riede Reference Riede2000). Remote sensing is generally regarded as being able to contribute to this aim, mainly through its ability to provide continuous spatial information (Leyequien et al. Reference Leyequien, Verrelst, Slot, Schaepman-Strub, Heitkönig and Skidmore2007). However, demographic parameters (e.g., age and sex structure) are needed for conservation decisions and appropriate management of animal populations. Previous studies show that the coefficient of variation of individual survey estimates of abundance often exceeded 50% (Seddon et al. Reference Seddon, Ismail, Shobrak, Ostrowski and Magin2003). Maximizing the extraction of available data can assist in getting more information from limited field data and improve overall data quality.
Historically, aerial imagery from crewed aircraft and a manual system where field biologists document visual observations represent an accepted method to estimate individual animal size (Koski et al. Reference Koski, Rugh, Punt and Zeh2006). High-resolution, low-altitude UAVs can be used to determine individual animal size, depending on the species and the survey environment (Watts et al. Reference Watts, Perry, Smith, Burgess, Wilkinson and Szantoi2010). Advances in photogrammetry software and the use of low-altitude imagery from UAVs (Berteška & Ruzgienė Reference Berteška and Ruzgienė2013) have transformed the way data are handled in a digital environment, which has added a range of possibilities to individual animal morphometric analysis.
UAV monitoring is not limited to the identification of species. Extracting additional information can give further insight, especially if enhanced by more nuanced perspectives of age structure, growth rates (Christiansen et al. Reference Christiansen, Vivier, Charlton, Ward, Amerson, Burnell and Bejder2018), sex ratios, reproductive status, body condition (Krause et al. Reference Krause, Hinke, Perryman, Goebel and LeRoi2017) and behaviour (Torres et al. Reference Torres, Nieukirk, Lemos and Chandler2018). This method provides the basis for rapidly and accurately measuring animal features from UAV data and meets a critical conservation need. Furthermore, this method is relatively non-invasive by nature (Horton et al. Reference Horton, Hauser, Cassel, Klaus, Fettermann and Key2019) and usually allows for the monitoring of species that can be difficult to physically measure otherwise (Durban et al. Reference Durban, Fearnbach, Barrett-Lennard, Perryman and Leroi2015). Remote data collection therefore allows for the collection of critical monitoring data without the increased stress of capturing and handling the animals.
UAV technology has shown great potential as a scientific monitoring tool; however, according to Jones et al. (Reference Jones, Pearlstine and Percival2006), this is only the case when combined it is with appropriate sensors, established sampling protocols and statistical analysis. The rapid growth in the use of UAV-acquired imagery for environmental monitoring (Laliberte et al. Reference Laliberte, Herrick, Rango and Winters2010, Gonzalez et al. Reference Gonzalez, Montes, Puig, Johnson, Mengersen and Gaston2016, Rey et al. Reference Rey, Volpi, Joost and Tuia2017) and the availability of off-the-shelf UAV units make this an attractive option for researchers. UAV environmental monitoring adoption by the conservation sector has lagged behind the technology sector, such that many technological possibilities remain underutilized. Applying low-altitude aerial imagery to conservation requires the coming together of three skill sets: ecology, UAV hardware and data interpretation (de Kock & Gallacher Reference de Kock and Gallacher2016).
This study was carried out to investigate whether UAV-acquired imagery could be used as a non-invasive monitoring tool for large ungulates in arid regions, and it focused on the Arabian oryx (Oryx leucoryx) as a pioneer species. The investigation was carried out to explore the mining of baseline data and to categorize the data in bracketed indexes to increase usability. The bracketed indexes were developed from predictive polynomial and decision tree models to estimate age groups, weights and confirmation of pregnancy within the herd when applied to geo-rectified UAV-acquired imagery.
The present study had two main objectives. Firstly, we examined the accuracy of species-specific zoometric measurements acquired by a non-invasive extraction from photogrammetrically processed drone-based imagery. Secondly, we tested the predictive value of UAV-acquired post-processed imagery to classify the herd structure in terms of age groups, including offspring identification, sex and pregnancy status of O. leucoryx (n = 43) in a protected area with the purpose to assist in the decision-making process for conservation management of the species.
Methods
Study species and sites
The Arabian oryx (O. leucoryx) is a large-bodied, arid-adapted antelope, and it is the only member within its genus that ranges outside of Africa. Compared with other species within the Oryx genus, its overall size makes O. leucoryx the smallest. The historical range of this medium-sized, desert-dwelling antelope covered the Arabian Peninsula and as far north as Syria (Harrison & Bates Reference Harrison and Bates1991). The species now occurs in a few reintroduced populations and many semi-captive populations in the Arabian Peninsula with substantially improved conservation status; yet most of the sites included in the 2011 assessment have had to be fenced (Mallon & Price Reference Mallon and Price2013), including our study site, the Dubai Desert Conservation Reserve (DDCR) in the United Arab Emirates (UAE).
The present research incorporated information on three herds. The first herd is referred to as the reference herd (n = 10); these animals were marked and data were collected from both a fixed camera and a UAV to compare the two types of digital image collection methods. The second herd, referred to as the study herd (n = 121), was used to collect baseline top-view imagery data (Supplementary Table S1, available online) where measurements and known data such as age, sex and weight were used to develop predictive models. The reference herd and study herd were both from managed captive populations. The third herd, referred to as the reintroduced herd, consisted of animals within the protected area. The imagery was collected from the study herd and the reference herd at Wadi al Safa Wildlife Centre (25.091200°N, 55.282360°E), the Arabian oryx conservation breeding centre, situated in the Emirate of Dubai, UAE (Fig. S1). A large managed O. leucoryx population is housed as part of the regional captive O. leucoryx conservation breeding programme. All individuals are handled yearly (O’Donovan & Bailey Reference O’Donovan and Bailey2006) for routine veterinary work, including vaccinations, health checks and breeding access and separation, as per best practice guidelines (de Kock et al. Reference de Kock, Al Qarqaz, Burns, Al Faqeer, Chege and Lloyd2018). Individual animal baseline data, including identifiers, weight, demographics and pregnancy status in females, were collected during this management process. Historically collected data such as date of birth were referenced from the zoo-based database.
Field data collected from digital measurements on captive O. leucoryx of the ‘study herd’ were compared with UAV-based survey data collected from the reintroduced herd in the DDCR (24.824789°N, 55.657069°E), a 225-km2 protected area that is home to over 800 reintroduced O. leucoryx to conserve the species. In order to validate the accuracy between the digital photo measurements and UAV imagery, the reference herd was measured using both techniques.
Validation image acquisition
As part of ongoing veterinary monitoring of captive and reintroduced oryx, both populations are captured annually and routine measurements, such as weight, are collected. During this process, we also validated photogrammetry techniques using a remotely triggered camera affixed to the animal management area during captures. Digitized morphometric measurements of the species were obtained using scale- and colour-rectified top-view imagery where the sex, age and weight of individual O. leucoryx were known, from the study herd (n = 121) and the reference herd (n = 10). This imagery was acquired by a remotely triggered camera mounted within the animal management area.
In order to calibrate these images, we placed vertical and horizontal scale bars in the separation areas, where the oryx were temporarily held for veterinary procedures (Fig. S2). The mounted scale bar was used as a point of calibration at the point of focus and permitted digital measurement of the total visible animal as seen from a top-view image; this image is referred to as the animal’s ‘drone print’, where the animal is present in the aerial view of UAV-derived imagery. Images of the animals taken in the separation area were recorded with either a GoPro 3 (GoPro, USA) action camera or with a Sony QX1 (Sony Cameras, Japan) with a Sony E 20 mm f/2.8 lens, both set to maximum resolution.
Digital zoometric measurements
In order to assess the accuracy of the digital measurements of the same individuals from a single scaled digital image and geo-rectified orthophoto-mosaic resulting from processed UAV-acquired imagery, 10 animals of the reference herd were marked for individual identification using a non-toxic, non-permanent coloured hairspray. A stencil with 12 squares in a grid of 4 × 3 (each square 7 cm × 7 cm) was cut into vinyl and placed on the back of the animals. Animals were marked consecutively as they were handled in a ‘tamer’ (Fauna Research, Red Hook, NY, USA), an animal restraining system used for captive ungulate management, and released into their holding pen. The markers were added to allow accurate identification in the comparison of the digital measurements with geo-rectified UAV imagery.
A range of limitations that includes the available number of animals, managing stress levels and environmental conditions during the handling process resulted in limited data. Because the animals were released into their enclosures as soon as possible after marking, some individuals were not visible during the subsequent UAV flight. With animal welfare as a prime concern, a second flight was not undertaken to avoid undue stress to the animals. As such, only six usable animal observations were utilized for the accuracy assessment.
Still images with reference scales of both the study and the reference herds were measured digitally. In the study herd, a total of 121 O. leucoryx were sampled (88 females and 33 males from the age of 1 month to 17 years). In the reference herd, 10 animals were measured (5 males and 5 females ranging from 3 to 14 years of age). A total of three measurements of each animal was collected: (1) total length from the base of the tail to the tip of the nose (cm); (2) overall two-dimensional area of the top view (cm2); and (3) width, namely the fullest part of the animal (cm) perpendicular to the dorsal median plain. All measurements were made from scaled imagery in two dimensions.
Individual top-view images were added to KLONK© Image Measurement Software Professional, Ver. 16.1.1.4 (Image Measurement Corporation, Cheyenne, WY, USA) for manual measurement and ImageJ©, Ver. 1.52a (Laboratory for Optical and Computational Instrumentation, University of Wisconsin, WI, USA) for semi-automated measurement using colour threshold analysis. The mounted scale bars in the separation area were used to rectify and scale the images on the focal length. Total body length (tip of the nose to base of the tail), and total visible area (top-view perspective) were measured for the individual animals. In addition to the digital measurements of the scaled photos of the study herd and the reference herd, their sex, weight, age and pregnancy status were recorded.
The reference herd (n = 10) was measured twice in different software applications. These double measurements were taken to compare the relationship between the manual digital scale imager measurements and the automated process resulting from the object-based image analysis (OBIA). Firstly, similar to the study herd, the images were scaled and measured using the same techniques. Secondly, automated measurements were taken from OBIA results in the form of exported polygon files of the drone print of each identified animal, using measurement tools within ArcMap 10.7.1 (ESRI, Redlands, CA, USA). The second method was also applied to the reintroduced O. leucoryx herd (n = 43) within the DDCR, which was used to validate the effectiveness of the predictive pregnancy model.
UAVs, sensors and control systems
Aerial photographs were taken using a DJI Inspire 2 UAV (DJI, China) with a Zemusse X5S camera. Flight planning was carried out with DroneDeploy v.2.0.11 (DroneDeploy, Santa Clara, CA, USA) software set to an above ground level (AGL) height of 100 m with a side overlap of 75% and a front overlap of 85%, as recommended for the photogrammetry software. The maximum speed of the UAV was set to 15 m/s. In order to increase the total spatial accuracy of the photogrammetry model, ground control points (GCPs) were collected and added to the photogrammetry processing to provide validation points for subsequent distance-based measurements of imagery. A total of five GCPs in the pattern of four points in the corners and one in the centre (de Kock & Gallacher Reference de Kock and Gallacher2016) of the surveyed area were added to the photogrammetry input. Each GCP’s geographical coordinates were acquired with <15 cm accuracy using a Trimble Geo 7X handheld GNSS (Trimble, Sunnyvale, CA, USA) system and added to the photogrammetry processing to improve the overall accuracy of the resulting orthophoto-mosaic.
During data acquisition of the reintroduced O. leucoryx in the DDCR, an Ebee Plus UAV (SenseFly, Switzerland) with a Canon S100 (Canon, Inc., Japan) camera was used. The flight plan was developed in Pix4Dcapture (Ver. 4.2.0) with a flight altitude of 100 m AGL and with picture intervals resulting in 80% side overlap and 80% front overlap, with an orthophoto-mosaic resolution of 4 cm/pixel.
Photogrammetric processing and analysis
The UAV-acquired and geotagged images with the required overlap needed for photogrammetric processing were processed using DroneDeploy online processing services (www.dronedeploy.com). The processed imagery was added to eCognition Developer © v.9 (Trimble, Sunnyvale, CA, USA) using an OBIA software (Blaschke Reference Blaschke2010), where the data were analysed using a developed ruleset (de Kock Reference de Kock2015) to extract O. leucoryx from the imagery using OBIA supervised classification. The extracted shapes were exported as a polygon shapefile that included data on the number of pixels within the polygon. The extracted size perimeters and digital measurements of marked individuals were compared to determine the accuracy of the drone imagery and secondly imported into R v3.6.2 (R Core Team 2013) statistical software.
The three available measurements from the processed images – total length, width (W) and total area (drone print) – were extracted in a semi-automated way. The width was rejected because of the possibility of an increase or decrease in these measurements over a relatively short period. Therefore, in animals that become fatter or thinner while maintaining the same or similar body length (such as in pregnancy where the drone print and width measurements will increase during pregnancy and decrease after birth), the total body length will be similar. Similarly, the body condition can be affected by environmental conditions such as droughts, where the abundance of food for the O. leucoryx is directly influenced. Normalizing the drone print using the total length of the animal allowed the data to be indexed with a single value, the normalized drone print index (NDPI):
$$NDPI = {{(DP - TL)} \over {(DP + TL)}}$$
The NDPI value range was 0–1, with DP (drone print) being the total top view of the animal (cm2) and TL (total length) being the distance from the tip of the nose to the base of the tail (cm). The NDPI values were categorized in a bracketed index that represents certain features. The application of the NDPI and the bracketed index on the DDCR dataset required the calculation of the NDPI in ArcMap. The OBIA extracted polygons representing individual animals were imported into ArcMap. Because the OBIA identified the objects from a raster image, the object boundary followed pixel boundaries, with the OBIA data consisting of a polygon shapefile and attribute data including the total number of pixels within the polygon and the end-to-end length of the polygon at its widest measurement. These two attribute values presented the DP and the TL of each extracted animal. The imported OBIA attributes included the pixel size and the number of pixels represented in the polygon; both values were used to calculate the DP (in cm2). The NDPI was calculated within ArcMap and added as an attribute value to the polygon shapefile.
Lastly, the selected range of the bracketing index was displayed to visualize the result in a geographical format.
Data analyses and model evaluation
The dataset of the reference herd was used to compare the physical animal zoometric measurements and the UAV-based, photogrammetry-processed imagery and digital measurements. The datasets from the study herd were used to develop predictive models. The dataset included all digital measurements as well as attribute data that included the date of birth, sex, weight, identification number and pregnancy status for females.
The DP data were normalized using the TL of the animals (Eq. (1)), resulting in a NDPI ranging from 0 to 1. The NDPI was used as the basis for the bracketing index, where a specific range of data was bracketed to present particular features. The average, mean average, minimum and maximum of each feature within the index were used as a guide to determine the best-suited bracket with the highest R2 and to include the noteworthy values of correlated data within the dataset.
A variety of models and model types was developed to investigate the best model fit, focusing on predictability. The inputs were used to develop the best-fitting model to predict the age, weight, sex and pregnancy of individual O. leucoryx. The model types included linear regression, polynomial and predictive models. Computer learning decision trees was also employed using R and the Classification and Regression Training (CARET) package (Kuhn Reference Kuhn2019).
The models included: the O. leucoryx weight predicting models; the two developed polynomial models using, respectively, the DP and NDP to predict weight; O. leucoryx age-predictive linear and polynomial models using DP and NDP; O. leucoryx pregnancy prediction based on the decision-tree models, using the DP, NDP, TL and W; and O. leucoryx sex prediction using the decision tree based on DP, NDP, TL and W data. During the polynomial model fit test, the Akaike information criterion (AIC) was used to reduce the potential of overfitting because the data used to develop the model were also used to test the overall fit.
Where the predicted data were known, the model fit was tested on the additional imagery; otherwise, the data were split into test and training subsets and the models were tested using the ‘test’ subset. The polynomial model fit was evaluated using the p-value, R2 and, where (x)2 was used, the adjusted R2.
The pregnancy status of the O. leucoryx in the DDCR was monitored by the rangers using researcher-led visual observations that were developed by the DDCR management team. The criteria included a visual assessment of physical features as well as behaviour, captured by well-experienced rangers.
Results
The accuracy was 98.41% with a standard error of the mean (SEM) of 0.401 for the DP and 97.99% with an SEM of 0.345 between the fixed camera image and zoometric measurements from processed UAV imagery of the reference herd (Table S2).
The candidate weight-predictive models consisted of two linear models and two polynomial models in order to determine which models were best supported by the data (Table S3). The linear regression model predicting the weight of individual O. leucoryx using the DP predicted weight produced a relatively high model fit (R2 = 0.799, overall comparison of AIC = 923.49). However, the model performed poorly based on the R2, especially considering residuals, which included younger animals that were growing rapidly to adult body size. The model using NDP (Fig. 1) resulted in an R2 of 0.7969 and an AIC of 948.83. Of the polynomial models using DP and DP2 and NDP and NDP2, the first resulted in an R2 of 0.85 and an AIC of 919.68, while the NDP and NDP2 model performed the best in terms of predicted weight (R2 = 0.85, AIC = 912.38).
Oryx leucoryx normalized drone print index (NDPI) and weight (n=121) graph, with a fitted polynomial regression model, with NDPI and normalized drone print index squared (DNPI2).
The age-predictive models using DP and NDP were not significant (polynomial regression using DP to predict age: R2 = 0.562, df = 119, residual stander error = 35.17; polynomial regression using NDP to predict age: R2 = 0.525, df = 119, residual standard error = 36.61). The models’ inability to predict age with relatively high precision is influenced by the growth curve that flattens out when antelope reach adulthood. The bracketed index, when focusing on age, showed more potential from birth to adulthood, possibly related to the natural breaks in the data and general lack thereof when the animals reach adult size.
The decision-tree models using width to predict pregnancy and the model using DP, NDP, TL and W had the same accuracy (0.917, p < 0.01). However, the model only using width was preferred because of the simplicity and the fact that the accuracy and p-value were not affected. The decision-tree model predicted that O. leucoryx with measurements >37 cm on the widest part of the drone print were pregnant. The high accuracy made this model a reliable tool to predict pregnancy in an O. leucoryx herd from UAV-based imagery.
The decision-tree model predicted that O. leucoryx with measurements ≥33 cm on the widest part of the drone print, NDPI >0.88 and total length >115 cm were males. The model fit test utilizing 30% of the data indicated 72.2% accuracy (p = 0.58). The model’s low accuracy, high p-value and low representation of adult male O. leucoryx within the training and test datasets negatively affected the usability of the predictive model.
Data with a high correlation with the DP allowed for predicted values using the NDPI to have a higher level of confidence. Data features such as the age of the animals that had a lower correlation with the DP were predicted to perform poorly when the NDPI was used to predict age in a population.
Applying the bracketed range of the NDPI (Table 1) to the sample data results allowed for the prediction of limited age group information, female pregnancy and weight ranges within the herd with high probability.
Bracketed index of Oryx leucoryx features with probabilities.
The bracketed NDPI was applied to a UAV-acquired and photogrammetry-processed orthophoto-mosaic of the reintroduced herd of O. leucoryx within the DDCR. The age bracket was applied to the reintroduced herd in the DDCR and showed 100% accuracy, as verified by trained rangers’ researcher-led visual observations.
The weight bracket could not be verified; verification would require the capture and weighing of these wild animals, and this was not an ethical or practical option. The pregnancy bracket index showed that 24 individuals were identified as possibly pregnant from the 43 individuals identified by OBIA; possibly pregnant females are shown on the map in Fig. S3 as green triangles. The DDCR rangers’ researcher-led observations suggested 21 females showed visible signs of pregnancy, resulting in an overall accuracy of the applied NDPI bracketed guideline for identifying a high probability of pregnancy in O. leucoryx of 87.5%. Lastly, the identification of females using the female bracketed index was 100% accurate.
Discussion
There was a remarkably strong relationship between zoometric measurements and the O. leucoryx pregnancy-predictive models. The developed models have a range of practical applications in the field of conservation. A field biologist responsible for the management of reintroduced O. leucoryx can use the tool to extract individual animal data applicable to the decision-making process. Furthermore, when multiple seasonal UAV-acquired datasets are used, these time-specific data comparisons may include calculations of the age structure, pregnancy and body condition scoring of the herd.
The combination of the pregnancy-predictive model and the age bracket to identify O. leucoryx aged <16 months, when used in yearly data analysis of the herd, can provide insights into the relationship between pregnancy and calf survival rates. The age-predicting model may improve if the data were filtered into a subset in a specific age range, with a better-fitted linear model.
However, there are limitations. This tool can only be used where the animals can be detected and extracted in a semi-automated or automated way with relatively high accuracy. Data from smaller animals will prove to be more challenging to extract and would be directly influenced by the resolution of the surveyed imagery. Vegetation, mist and other environmental conditions limit image capture of the species of interest with traditional red, green and blue (RGB) sensors. Similarities in the relationship between the animal coat colour and the environmental background may be challenging to the detection process, especially if the survey data are limited to RGB camera sensors. The errors may be influenced by a range of factors that can include the quality of the data, the overall accuracy of the images processed by photogrammetry software resulting in a geo-rectified orthophoto-mosaic, the cm/pixel size of the resulting orthophoto-mosaic, the sensor used and movement of the animals during the aerial survey. Further research in the area of animal detectability, detection probability and the limitations of using OBIA is needed.
Safe operation of the UAVs is the primary concern of drone pilots, and meteorological conditions may limit the performance of the UAV and result in grounding of the plane. The survey area is usually limited to UAV operations and, in most cases, the flight time per battery (Zhang et al. Reference Zhang, Hu, Lian, Fan, Ouyang and Ye2016). Better sensors with increased resolution are a current trend. These result in a larger file size for each picture and therefore an increase in the dataset size, putting more pressure on computer hardware and increasing the cost of data analysis (Casella et al. Reference Casella, Collin, Harris, Ferse, Bejarano and Parravicini2017).
UAV legislation (Luppicini & So Reference Luppicini and So2016, Stöcker et al. Reference Stöcker, Bennett, Nex, Gerke and Zevenbergen2017) has been created in some countries and developed in others over the last decade. In a range of countries, UAVs are limited to dedicated or shared airspace, with formal legislation limiting altitude, area of operation, UAV type and formal UAV pilot training, thus making use of UAVs for field observations in conservation a complicated task. Field ecologist skillsets may be limited for deploying this tool effectively; a combination of UAV operations, photogrammetry software and geographical information system (GIS)-based data analysis skills are currently required to mine UAV data effectively. Future challenges include the automation of the zoometric data extraction process in an interface that is user-friendly, with technical analysis running in the background, as well as providing an answer for the end user, usable results and results of value for the field biologist managing endangered species.
Ground-based monitoring of O. leucoryx in protected areas is commonly used as an observation and survey technique at ground level, and UAV surveys provide an efficient aerial alternative to some of these traditional survey techniques. The mostly desert terrain in the historical range of the species is relatively difficult to navigate; UAV surveys and the ability to extract additional animal-specific information may make this an attractive tool for future use.
Conclusions
Zafar-ul Islam et al. (Reference Zafar-ul Islam, Ismail and Boug2011) have suggested that, historically, during post-release monitoring of O. leucoryx, the low-density distribution of these reintroduced animals resulted in a population size estimate with low accuracy. The DDCR, with a single flagship species, is dedicated in its efforts to protect O. leucoryx; however, there is a need to continuously monitor herd health, without the risk of adverse negative effects to the re-wilding strategy. UAV-acquired imagery proved to be an effective tool and can assist in providing critical information to reserve management in a non-invasive manner.
Future studies should aim to develop practical tools to support decision-making and focus on management-orientated results to assist in species and population management, especially in a delicate arid environment. Future studies on UAV-based zoometric data analysis are suggested to investigate the extraction of more features, such as animal height, which has practical application in large mammal age classification. Furthermore, animal height can be combined with the DP to calculate animal volume, and this may assist in individual animal body condition predictions from UAV imagery. This information will give protected area managers a better insight into the physical condition of individual animals and the overall herd, especially during high-risk periods that may include those after reintroductions, disease outbreaks and droughts. Moreover, the advantage of the imagery data is that they can be reprocessed and analysed in the future if needed, or if advances in analysis and software allow for more comprehensive data mining.
Supplementary Material
To view supplementary material for this article, please visit https://doi.org/10.1017/S0376892921000242.
Acknowledgements
We gratefully acknowledge firstly His Highness Sh Hamdan Bin Rashid Al Maktoum, Deputy Ruler of Dubai and Minister for Finance of the UAE, all the staff of Wadi Al Safa Wildlife Centre for their help and support during the data acquisition phase of this study and the management staff of the Dubai Desert Conservation Reserve (DDCR, Dubai, UAE).
Financial support
Financial support for this research was provided by the Faculty of Tropical AgriSciences at the Czech University of Life Sciences Prague, projects CIGA 20185008 and IGA 20213103.
Conflict of interest
None.
Ethical standards
The authors assert that all procedures contributing to this work comply with applicable national and institutional ethical guidelines on the care and use of laboratory or otherwise regulated animals.
