Application of UAV Remote Sensing for a Population Census of Large Wild Herbivores—Taking the Headwater Region of the Yellow River as an Example

: We used unmanned aerial vehicles (UAVs) to carry out a relatively complete population census of large wild herbivores in Maduo County on the Tibetan Plateau in the spring of 2017. The effective area covered by aerial surveys was 326.6 km 2 , and 23,784 images were acquired. Interpretation tag libraries for UAV images were created for wild animals, including Kiang ( Equus kiang ), Tibetan gazelle ( Procapra picticaudata ), and blue sheep ( Pseudois nayaur ), as well as livestock, including yaks and Tibetan sheep. Large wild herbivores in the survey transect were identiﬁed through manual imagery interpretation. Densities ranged from 1.15/km 2 for Kiang, 0.61/km 2 for Tibetan gazelle, 0.62/km 2 for blue sheep, 4.12/km 2 for domestic yak, and 7.34/km 2 for domestic sheep. A method based on meadows in the cold and warm seasons was used for estimating the densities and numbers of large wild herbivores and livestock, and was veriﬁed against records of livestock numbers. Population estimates for Kiang, Tibetan gazelle, blue sheep, domestic yak, and domestic sheep were 17,109, 15,961, 9324, 70,846, and 102,194, respectively. Based on published consumption estimates, the results suggest that domestic stock consume 4.5 times the amount of vegetation of large wild herbivores. Compared with traditional ground survey methods, performance of UAV remote sensing surveys of large wild herbivore populations was fast, economical and reliable, providing an effective future method for surveying wild animals. survey transect overlapped with the drone survey line and the time difference between the two surveys was 1 h. Three Kiang groups were found by ground survey, with counts of 31, 51, and 30 (total count = 112). In the UAV survey, three groups of Kiang were found, with counts of 28, 48, and 31 (total count = 107). The three Kiang groups that were found with the two survey methods in the same place deviated in quantity by 9.68%, 5.88%, and 3.33%, with an average deviation of 4.46%. On 15 and 16 April the time differences between the two surveys were <1 h. No animals were found on either the ground or drone surveys, so the results were consistent.


Introduction
Terrestrial wild animal populations can be investigated with ground surveys, aerial surveys  [25]. Therefore, a census of the large wild herbivore population in the county is urgently needed for informing wildlife and grassland management. At the same time, a study of the population distribution could provide important information for the delineation of protected areas and reasonable grazing by herders.

Materials and Methods
In this study, we conducted UAV remote sensing to survey the population and distribution of large wild herbivores in Maduo County, in which the Yellow River Source Park of the TRS National Park is located. The survey method is shown in Figure 1.
Remote Sens. 2018, 10, x FOR PEER REVIEW 3 of 15 and protection in the TRS area, the TRS National Park was established in 2016 as China's first national 96 park, incorporating 78.1% of Maduo County's land area [25]. Therefore, a census of the large wild 97 herbivore population in the county is urgently needed for informing wildlife and grassland 98 management. At the same time, a study of the population distribution could provide important 99 information for the delineation of protected areas and reasonable grazing by herders.

Study Area
Maduo County is located at the source area of the Yellow River and is the main production region in the area. The county is densely populated with rivers and lakes and 78.1% of the county's land is within the Yellow River Source Park of Sanjiangyuan National Park. The county is rich in resources, containing more than 50 species of wild vertebrate, including Kiang (Equus kiang), Tibetan gazelle (Procapra picticaudata), and blue sheep (Pseudois nayaur). Maduo County is attached to the Guoluo Tibetan Autonomous Prefecture, located in the south of Qinghai Province, at the northern foot of Bayankala Mountain (96 • 50 −99 • 20 east longitude, 33 • 50 −35 • 40 north latitude). The county is 207 km from north to south and 228 km from east to west, with a total land area of 25,300 km 2 , an altitude range of 3902-5243 m, and an average elevation of 4200 m. The area consists mainly of high plains, with little relief, and is relatively flat and high in the southwest, and low in the northeast, containing flatlands, deserts, and marshes. The vegetation of Maduo County is mainly grassland, which accounts for approximately 88% of the area, and includes alpine grassland and alpine meadow.

UAV Survey
In accordance with the National Terrestrial Wild Animals Resources Survey and Monitoring Technical Specification from the SFA, we used a sampling intensity of "not less than 2.0% for grassland areas and not less than 1.0% for meadow areas." The proportion of different terrains, land use/cover, and vegetation types was integrated, and systematic sampling was used to develop the survey area and flight plan ( Figure 2). An aerial survey was carried out on 9-18 April 2017 from 8:00 to 11:00 a.m., when the solar elevation angle was low, and the surveyed individuals cast clear shadows on the ground, which reduces interpretation difficulty and improves accuracy. We used two electric fixed-wing aircraft from the Chengdu Institute of Mountain Hazards and Environment at the Chinese Academy of Sciences ( Figure 3a) and the Shenzhen FEIMA Robotics Technology Co., Ltd, Beijing, China ( Figure 3b). The UAV system and image parameters are shown in Table 1. We conducted 14 samples with an effective aerial area of 326.6 km 2 with an aerial image resolution of 4-7 cm, providing a total of 23,784 images. The samples covered an area of 72.85 km 2 in alpine grassland with a sampling rate of 2.21%, and an area of 249.69 km 2 in alpine meadow with a sampling rate of 1.42%.

122
Technical Specification from the SFA, we used a sampling intensity of "not less than 2.0% for 123 grassland areas and not less than 1.0% for meadow areas." The proportion of different terrains, land 124 use/cover, and vegetation types was integrated, and systematic sampling was used to develop the 125 survey area and flight plan ( Figure 2). An aerial survey was carried out on 9-18 April 2017 from 8:00

Image Recognition and Quality Control
Two high-performance workstations were used for mosaicking images, and five computers were used for visual interpretation. Image mosaics were constructed using Pix4Dmapper, FeiMa, and LiMapper; visual interpretation was performed in ArcGIS.
For our visual interpretations, we established applicable tag libraries according to seven elements of remote sensing interpretation, including tone, color, texture, shadow, size, shape, and layout ( Table 2). The main body is smoke brown, ochre brown, with edges in white and brown black.
The main body is earthy yellow, yellowish brown and grayish yellow, with tail end in white.
Body can be steel gray, gray and dirty white.
The main body is black, gray black and white.
The main body is black, dirty white and white.

Texture
The brown longitudinal strips in the center of back, and splicing texture of white patches at edges of the limbs and body side.
It is earthy yellow or solid color of similar color, with white patches on hips.
Gradients from steel gray to gray white. Juvenile sheep can be as small as 0.4 m, but will not be alone.

Shape
The overall shape is approximate to the long strip and bulk or long handle and long circle. Length-width ratio is 4:1-5:1.
The overall shape is approximate to the long oval or rod-like. Length-width ratio is 3:1-5:1.
The overall shape is approximate to the long oval or rectangle. Length-width ratio is 1.5:1-3:1.
The overall shape is approximate to the long oval or water drop-shaped. Length-width ratio is 1.5:1-3:1.

Layout
Mostly distribute in groups, a small number of individuals live alone.
Live in Group or alone. Mostly live in groups.
Mostly distribute in groups, often with objects and traces of human activities in the neighborhood.
Mostly distribute in groups, often with objects and traces of human activities in the neighborhood.

Ground Surveys
Ground surveys were conducted on foot by four observers, matching the UAV survey transects and using a HCIYET HT-1500A rangefinder and compass. The following variables were recorded: wild animal species, number of individuals in the groups, geographical coordinates of the record points, distances between record points and wild animals, angle between the observation direction and survey route, survey time, and route length. A direct counting method was applied for the ground survey. To avoid visual errors that increase with distance, only areas within 500 m on both sides of the survey route were recorded. Individuals were defined as being part of the same cluster if the distance between them was <100 m.

Population Estimation
The transect densities of wild and domestic herbivores were calculated using the UAV image interpretation results, as well as five analysis methods, including direct extrapolation, extrapolation after deducting non-vegetated areas, and estimations based on meadows in the cold and warm seasons, on elevation zones, and on vegetation type. Domesticated herbivores were also estimated, and the results were verified using livestock number records provided by the Qinghai Provincial Grassland Station, to identify the most accurate and reasonable method.
Through direct extrapolation, populations were estimated by assuming the animal density in the transects to be the same as the density in Maduo County. The other four methods were estimated based on calculating animal populations in different areas with various properties, and then adding the numbers using the following formula: where Q represents the estimation of the population of any kind of animal in Maduo County; D Si represents the density of animals in transects with properties i; A i represents the total area with properties i, and n is the number of properties. Data for seasonal meadows were provided by the Qinghai Provincial Department of Agriculture and Animal Husbandry, and the division of seasonal grazing lands was mostly based on the range of suitable grazing in different seasons formed by long-term grazing activities.

Population and Group Size in the UAV Survey Transects (1) Populations
The sheep unit refers to the unit of livestock calculation, and was converted according to the animal's food intake compared with adult domestic sheep. To objectively reflect the relationships between surveyed objects and the surrounding grassland, we converted the surveyed subjects into sheep units, as shown in Table 3.  Our survey statistics revealed that, of the 14 transects surveyed by UAV in spring 2017, the greatest number of large herbivores (1043) was found in transect 11, accounting for 30.8% of the total. We found 195 blue sheep in transect 11, accounting for 96.06% of the blue sheep in all transects. The highest number of Kiangs (107) was found in transect 1, accounting for 28.31% of the total. The highest number of Tibetan gazelles (72) was found in transect 12, accounting for 36.18% of the total. The highest number of domestic yaks was found in transect 1, accounting for 47.3% of the total, while domestic sheep were most often found in transect 9 (n = 891), accounting for 37.04% of the total (Figure 4).  Our survey statistics revealed that, of the 14 transects surveyed by UAV in spring 2017, the greatest number of large herbivores (1043) was found in transect 11, accounting for 30.8% of the total. We found 195 blue sheep in transect 11, accounting for 96.06% of the blue sheep in all transects. The highest number of Kiangs (107) was found in transect 1, accounting for 28.31% of the total. The highest number of Tibetan gazelles (72) was found in transect 12, accounting for 36.18% of the total. The highest number of domestic yaks was found in transect 1, accounting for 47.3% of the total, while domestic sheep were most often found in transect 9 (n = 891), accounting for 37.04% of the total (Figure 4). (2) Accuracy verification Based on the 2017 synchronous ground observation data from 9, 15 and 16 April, we conducted a comparative analysis of the credibility of our UAV image interpretations. On 9 April, the ground (2) Accuracy verification Based on the 2017 synchronous ground observation data from 9, 15 and 16 April, we conducted a comparative analysis of the credibility of our UAV image interpretations. On 9 April, the ground survey transect overlapped with the drone survey line and the time difference between the two surveys was 1 h. Three Kiang groups were found by ground survey, with counts of 31, 51, and 30 (total count = 112). In the UAV survey, three groups of Kiang were found, with counts of 28, 48, and 31 (total count = 107). The three Kiang groups that were found with the two survey methods in the same place deviated in quantity by 9.68%, 5.88%, and 3.33%, with an average deviation of 4.46%. On 15 and 16 April the time differences between the two surveys were <1 h. No animals were found on either the ground or drone surveys, so the results were consistent.
We compared the aerial image interpretation results of the two drone types for the same region, and we found that the interpretations of the two resolutions both identified 28 Kiang, yielding 100% agreement. (Figure 5) Remote Sens. 2018, 10, x FOR PEER REVIEW 10 of 15 survey transect overlapped with the drone survey line and the time difference between the two surveys was 1 h. Three Kiang groups were found by ground survey, with counts of 31, 51, and 30 (total count = 112). In the UAV survey, three groups of Kiang were found, with counts of 28, 48, and 31 (total count = 107). The three Kiang groups that were found with the two survey methods in the same place deviated in quantity by 9.68%, 5.88%, and 3.33%, with an average deviation of 4.46%. On 15 and 16 April the time differences between the two surveys were <1 h. No animals were found on either the ground or drone surveys, so the results were consistent. We compared the aerial image interpretation results of the two drone types for the same region, and we found that the interpretations of the two resolutions both identified 28 Kiang, yielding 100% agreement. (Figure 5) Five methods, including direct extrapolation based on land area, extrapolation after deducting non-vegetated areas, and estimations based on meadows in cold and warm seasons, elevation zones, and vegetation type, were used to calculate the total number of domestic herbivores, and were compared with livestock number records provided by the Qinghai Provincial Grassland Station.
According to those records, at the end of 2015, there were 59,235 domestic yak and 73,133 domestic sheep in Maduo County. Because the records were from the end of 2015, we converted the UAV survey results from spring 2017 to the number of livestock at the end of 2016, according to a 30% birth rate. Because we assumed that the number of livestock in 2016 was the same as in 2015, the number of livestock converted from UAV data could be compared with the livestock number records at the end of 2015 (Table 5).  Five methods, including direct extrapolation based on land area, extrapolation after deducting non-vegetated areas, and estimations based on meadows in cold and warm seasons, elevation zones, and vegetation type, were used to calculate the total number of domestic herbivores, and were compared with livestock number records provided by the Qinghai Provincial Grassland Station.
According to those records, at the end of 2015, there were 59,235 domestic yak and 73,133 domestic sheep in Maduo County. Because the records were from the end of 2015, we converted the UAV survey results from spring 2017 to the number of livestock at the end of 2016, according to a 30% birth rate. Because we assumed that the number of livestock in 2016 was the same as in 2015, the number of livestock converted from UAV data could be compared with the livestock number records at the end of 2015 (Table 5).

Discussion
Given the constraints of manpower and material conditions, the intensity of wildlife census by ground survey is around 1% [26], and the general known ground-survey research area in the Qinghai-Tibet area is less than 7,500 km 2 [6,27]. We used UAV remote sensing for a large wild herbivore census in spring 2017, and we targeted a 2.21% sampling rate in alpine grassland and 1.42% in alpine meadow, over a 25,300-km 2 study area. While our sampling intensity was significantly higher than that of traditional ground surveys, the study area increased more than three times. The Tibetan Plateau environment is complex, and some wild animal habitats are difficult to reach. Ground surveys can only drive vehicles along the road, so it is impossible to investigate species that have special habitat requirements, such as blue sheep. However, UAVs are not limited by the ground environment, and the layout of sample plots is more consistent with statistical requirements, and estimation results are more accurate and reliable. Furthermore, UAVs are small and they do not make much noise, which eases the problem of wildlife sensitivity to disturbance. It is worth noting that UAV surveys are top-down, and are difficult to carry out in areas that are severely shaded by trees; but this is rarely a problem in the Tibetan Plateau because of its exposed natural environment.
We accumulated some useful experiences and developed a complete set of technical methods that can be used in future research. We used UAV aerial images that were 4-7 cm, and Kiang, Tibetan gazelle, blue sheep, domestic yak, and domestic sheep were clearly identifiable. Therefore, we recommend not exceeding 7 cm for image resolution in future UAV surveys of large wild herbivores, with an optimal resolution of at least 5 cm. There was an 80% front overlap and a 60% side overlap in the UAV survey, but problems with low accuracy of image mosaics were still present because of fewer feature points in the study area. These findings suggest that, in future research on the Tibetan plateau or similar areas, further image-overlap enhancement will have a limited effect on improving mosaic accuracy, and methods such as control-point placement on the ground should be used.
Herein, we constructed an image interpretation tag library for UAV remote sensing of large herbivores, based on Kiang, Tibetan gazelle, blue sheep, domestic yaks, and domestic sheep. In future, the objects included in the remote sensing image interpretation tag library for UAVs could be continuously expanded to important wildlife other than large herbivores in the Tibetan Plateau, such as snow leopards (Uncia). Additionally, this library could provide technical support for wildlife monitoring in the Tibetan Plateau. Moreover, there are a growing number of studies using automated or semi-automated detection of wild animals, including a study by Norouzzadeh et al., wherein use of camera-trap images with deep learning to conduct automatic identification, counting, and description of wild animals achieves 96.6% accuracy [28]. Pabico et al. identified animal breeds and species automatically using bioacoustics and artificial neural networks [29]. Automatic identification of wild animals based on UAV remote sensing images could also be developed to reduce manual operation or replace the image recognition work of visual interactions.
Because of a lack of comparable wildlife data, we used indirect verification of livestock number records. However, the most accurate method for domestic species is not necessarily the most suitable method for wild herbivores, and may have caused extrapolation errors. The 2017 survey results we used in this study revealed densities of 0.77/km 2 for Kiang and 0.72/km 2 for Tibetan gazelle in Maduo County. The results of the first national terrestrial wild animals resource survey conducted by the SFA from 1995 to 2003 reported densities of 0.874/km 2 for Kiang and 0.557/km 2 for Tibetan gazelle in the Qinghai province [30]. There is insufficient evidence for a significant increase in the number of wild animals, indicating that wildlife in this area require ongoing protection.
Although many researchers have examined the feeding intake and preferences of wild herbivores [2,[30][31][32], research on forage-livestock balance has often focused on four aspects: grassland forage yield, livestock feed intake, pasture availability, and stock capacity calculations [33][34][35][36][37][38][39][40][41][42]. Thus, these studies have not examined the effects of large wild herbivores on the forage-livestock balance. Meadows that are not categorized as over-grazed by only measuring livestock may actually be over-grazed because of the presence of large wild herbivores. Our findings suggest that, for ensuring adequate food and other living resources, the core area of the TRS National Park requires further reduction of livestock to protect wildlife.
UAV survey methods can be used to obtain annual and inter-annual data on wild animal populations in the TRS National Park over a long time period, monitor changes in wild animal populations more accurately, and quantitatively assess the effectiveness of biodiversity conservation in the TRS National Park. These findings can provide the basis for further formulation and implementation of protection measures.

Conclusions
In transects surveyed by UAVs in Maduo County in spring 2017, the transect density was 1.15/km 2 for Kiang, 0.61/km 2 for Tibetan gazelle, 0.62/km 2 for blue sheep, 4.12/km 2 for domestic yak, and 7.34/km 2 for domestic Tibetan sheep. In the spring, we found that a small number of Kiang and Tibetan gazelle lived alone, while the majority lived in groups. Blue sheep were not found to live alone, and generally lived in groups of less than 50. Our estimations revealed that Maduo County contained  17,109 Kiangs, 15,961 Tibetan gazelles, 9324 blue sheep, 70,846 domestic yaks, and 102,194 domestic sheep. The ratio of large wild herbivores to livestock was 1:4.5 in sheep units.
We found no significant increase in the population of large wild herbivores, indicating that further protection is still needed. Our findings suggest that the core area of the TRS National Park requires a further reduction in the number of livestock to ensure that sufficient resources are available for wildlife.