Vegetation Type Preferences in Red Deer (Cervus elaphus) Determined by Object Detection Models

Abstract

This study investigates the possibility of utilising a drone equipped with a thermal camera to monitor the spatial distribution of red deer (Cervus elaphus) and to determine their behavioural patterns, as well as preferences for vegetation types in a moor in Denmark. The spatial distribution of red deer was mapped according to time of day and vegetation types. Reed deer were separated manually from fallow deer (Dama dama) due to varying footage quality. Automated object detection from thermal camera footage was used to identification of two behaviours, “Eating” and “Lying”, enabling insights into the behavioural patterns of red deer in different vegetation types. The results showed a migration of red deer from the moors to agricultural fields during the night. The higher proportion of time spent eating in agricultural grass fields compared to two natural vegetation types, “Grey dune” and “Decalcified fixed dune”, indicates that fields are important foraging habitats for red deer. The red deer populations were observed significantly later on grass fields compared to the natural vegetation types. This may be due to human disturbance or lack of randomisation of the flight time with the drone. Further studies are suggested across different seasons as well as the time of day for a better understanding of the annual and diurnal foraging patterns of red deer.

1. Introduction

1.1. Monitoring Possibilities in Population and Conservation Biology

Some of the most important factors for successful species management and conservation are reliable tools for monitoring species and their population size and reproduction [1]. Information on the number of animals that have been killed or sighted during hunting is often used as an indirect measure for estimating the population size. However, yearly game bag records are often biased and may sometimes reflect the traditions of hunting more than the demographic changes in a population [2,3]. Alternatively, direct counting methods, such as faecal density counts, images from camera traps, and ground-based transect surveys, are employed. These methods are time–consuming, and potential disturbance of the animals could affect the results and present no details of foraging behaviour. Recently, the use of aerial drones has been increasingly used in the monitoring of population sizes and dynamics [4,5]. The combination of drones and thermal cameras makes monitoring of nocturnal species and other species that may be challenging to detect easier and more precise. Monitoring of nocturnal species like red deer (Cervus elaphus) has traditionally been executed by nightly spotlight surveys [6,7,8,9]. The advantages of using drones with thermal cameras are many, including a higher detection rate of nocturnal species, higher spatial coverage than traditional methods, the opportunity to examine behaviour, higher time efficiency, and lower disturbance of the animals [5,10,11,12,13,14,15]. The use of drones compared to traditional methods has been studied on several occasions, where the results showed that drones were more efficient and accurate [13,16,17,18]. The applicability of drones for studying and monitoring wildlife is extensive in terms of species and habitat types, as recent studies demonstrated the applicability of drones for differentiating between species as well as individuals [19,20,21]. The detection of individuals and estimation of behaviour has mainly been restricted to studies conducted on production animals [22,23,24]. However, the enabling of the detection of individuals and estimation of their location, age, sex, and especially behaviour entails a widening of the range of possible studies. More knowledge about the foraging behaviour and nutritional demands of deer may be obtained by combining the thermal drone data with information on present vegetation types, habitat composition, anthropological infrastructure such as roads, fields with crops, human disturbance, and crop types. This combination enables analysis of how and when animals use their habitats, which is especially relevant for determining and preventing agricultural damage, particularly caused by red deer.

1.2. Ecology and Biology of Red Deer

The red deer population in Denmark is estimated to be approximately 33,000 individuals distributed throughout most of the Jutland peninsula [25,26]. The hunting season is known to affect the home range of red deer, as red deer are highly sensitive to disturbances such as noise. The home range of a red deer individual in Denmark is estimated to be around 500 hectares outside the hunting season, but their activity increases during the hunting season, expanding their range to approximately 1100 hectares [27]. According to a report from Aarhus University on game yield reports, 536 individuals of red deer were reported by hunters in Thisted municipality in the hunting season 2022/2023, where this study was conducted [28]. The density of the game yield was 0.49 individuals km² in Thisted municipality, making it one of the highest in Denmark (see Figure A1) [28]. Red deer are both intermediate feeders and concentrate selectors [29,30]. Moreover, other important food resources in Denmark are spruce and heather [25,30]. The diet of red deer in moor vegetation includes the dwarf shrub Calluna vulgaris, as well as the genus Vaccinium which consists of dwarf bushes and small trees [29]. It is estimated that the average daily dry matter intake of red deer is approximately 1.5 kg [31]. Red deer are primarily found in forests and plantations in Denmark. They often forage in open landscapes and agricultural fields. Foraging mainly occurs after sundown and before sunrise, as disturbances from human activities are often less frequent during the night [25]. Red deer are gregarious animals found in mixed herds or herds of the same sex. Herds are often mixed during the heat period in September. Afterwards, the sex-separated herds are once again formed, dividing males (stags) and females (does).

1.3. Agricultural Damage Associated with Red Deer

Crop raiding red deer may cause costly yield losses and conflicts between agriculture, game management, and conservation. Damage that red deer are known to cause includes animals removing valuable crops, peeling the bark of trees, lying in crops, opening forage pits, treading on crops, and destroying fences [32]. The reason for red deer damaging the crops and agricultural structures may be the lack of natural quality habitats with sufficient nutritional content to support the local population of red deer [32]. According to a study by Corgatelli et al. (2019), the amount of damage to crops of maize resulting from direct foraging is related to the density of red deer, as difference in biomass production pr. maize plant between grazed and non-grazed plots in areas with low densities of red deer (0–1.6 heads/km²) was lower than in areas with high densities of red deer (14–30 heads/km²) [33]. The relationship between the amount of damage to agricultural crops and the density of deer is non-linear, and there seems to be a threshold density before which no damage occurs [34,35,36]. A study by Borkowski et al. (2019) showed that the use of forest areas by red deer is strongly affected by the presence of surrounding non-forest sites [37]. They found that the density of red deer was highest in the proximity to non-forest sites providing food, such as arable lands and meadows [37]. The selection of forage sites in ungulates is a trade-off between maximising nutritional abundance while minimising risks related to predation, hunting, and human disturbance [38]. Damage to crops and agricultural facilities caused by red deer is of a significant magnitude and as red deer are nocturnal foragers, it necessitates other monitoring solutions. Therefore, the development of methods exploiting the novel advancements in drone technology and object detection models combined with thermal cameras for studying their forage behaviour is of great interest and importance to preservation organisations and farmers alike. This study aims to investigate the foraging behaviour of red deer utilising a drone equipped with a thermal imaging camera in a moor in Northern Jutland, Denmark and gain knowledge of the following:

• The distribution of red deer in time and space in natural vegetation types and agricultural fields;
• The types of behaviour of red deer using automated object detection from thermal camera footage;
• The vegetation type preference of red deer;
• The behavioural patterns exhibited by red deer in different vegetation types.

2. Materials and Methods

2.1. Image Collection and Study Areas

This survey took place within and near Lyngby Hede, Thy, Denmark. The study was conducted for four nights (07:00 p.m.–06:00 a.m.), 4–5 March, 12–13 March, 14–15 March, and 15–16 March, of 2024 with an average temperature of 5.2 °C and maximum wind gust up to 24.6 m/s. A flight could be cancelled due to adverse weather conditions, specifically when strong winds made it difficult to control the drone and when rain or fog impaired camera visibility. Each night consisted of 6–8 flights of 15–30 min each, resulting in 1.5–4 h of observation per night. This study covered approximately 34.5 km.

2.2. Data Collection

To investigate the foraging behaviour and vegetation type preferences of red deer, this study utilised a DJI Matrice 300 RTK (M300 UAS) drone (SZ DJI Technology Co., Nanshan, Shenzhen, China) drone equipped with a high-resolution thermal camera, Zenmuse H20N (SZ DJI Technology Co., Nanshan, Shenzhen, Kina). This camera enables a thermal image resolution of 640 × 512 and up to 32× zoom. Video for data collection was recorded with a frame rate of 30 frames per second and was subsequently compressed to one frame per second. Since each second was represented by one frame, the resolution of behaviour was one second. The total number of frames used for behaviour prediction was approximately 12,000 frames. The drone was also equipped with a pinpoint marker function integrated into Google Maps, allowing the determination of the location of individuals and populations. The drone was piloted manually during the surveys and was exclusively flown in allowed areas with a maximum altitude of 120 m above takeoff altitude following legislation and a minimum altitude of 70 m above takeoff altitude to reduce disturbance of animals. The scouring drone method developed by Povlsen et al. (2023) was used [18]. This method entails that the drone was manually flown, and deer populations were spotted at a high flight height and thereafter inspected at a lower flight height and a distance. Population counts were conducted by positioning the drone at a long distance from the target to encompass the entire population within the frame. Subsequently, as the drone was brought closer, it facilitated the identification of species and behaviours for further analysis. Finally, the drone was brought back near the takeoff site before a new direction was explored. Multiple vegetation types were covered within and near Lyngby Hede, including moor, fields and to a limited extent, forest, providing a comprehensive understanding of the red deer’s spatial distribution and habitat use patterns. Data on vegetation cover were acquired from the Danish National Environmental Portal [39].

2.3. Species Identification

Several key morphological features were used to differentiate between red deer and fallow deer (Dama dama). Foremost among these was the subjectively evaluated body size, with red deer having the largest bodies (adult males ranging from 100 to 220 kg pre- and post-rut average and females ranging from 55 to 130 kg) and fallow deer the smallest (adult males weigh ~100 kg and adult females weigh ∼50 kg) [40]. Additionally, facial morphology provided additional characteristics for distinguishing between the two species of deer, with red deer displaying elongated and broad faces, while fallow deer have relatively large, yet shorter, facial structures (see Figure 1a,b) [1]. Moreover, the red deer demonstrates longer ears in comparison to the fallow deer (see Figure 1c), and the latter occasionally displays white markings around the neck and backside (see Figure 1d). In instances where direct assessment of body size and facial features was hindered by sub-optimal camera angles, behavioural cues were utilised for species identification. Specifically, differences in locomotion patterns were utilised, where red deer typically demonstrate a powerful and robust gait due to their larger physique, whereas fallow deer are more slender, resulting in comparatively agile movements. Consequently, red deer demonstrated longer and slower steps, contrasting with the shorter and faster steps observed in fallow deer.

2.4. Behaviour Identification of Red Deer and Fallow Deer

The recorded videos were processed and annotated using Roboflow (version 1.0), and behaviour classification was performed using the YOLOv8 (You Only Look Once) algorithm by Ultralytics (Los Angeles, CA, USA) [41,42]. Two video recordings, one lasting 16 m and 14 s and the other 20 m and 16 s, featuring herds of red deer showcasing all described behavioural classes, were uploaded to Roboflow as training data for the algorithm. These videos were uploaded with one frame per second, resulting in 2190 frames, and were subsequently annotated. Annotation of the frames enabled the identification and categorisation of various behavioural patterns exhibited by red deer. The classes of behaviours included “Eating”, “Standing”, “Lying”, “Locomotion”, “Heat spot”, “Multiple together”, and “Other”, and the algorithm was trained on footage of both red and fallow deer (see

Table 1. Ethogram for red deer behaviours.

Behaviour	Description
Eating	Animals standing with their head lowered towards the ground
Standing	Animals standing up and surveying their surroundings
Lying	Animals lying down
Locomotion	Animals being in motion, either running or walking
Heat spots	Spots on the ground with elevated temperatures caused by deer lying
Multiple together	Multiple animals being close in proximity hindering separation
Other	E.g., scratching itself or instances where behaviour was inconclusive

). Additionally, other species were observed in the study area, and therefore, two additional classes were added, “Horse” and “Sheep”. These were added with the intention of increasing the accuracy of the models predictions of the behaviours of red deer.

The classification of the 2190 frames resulted in a dataset of 808 frames with multiple animals and classes of behaviour on each frame. The missing 1382 frames were null frames that were removed, resulting in a dataset containing 30% of the null annotated frames. The annotation mode was “object detection”, and ∼500 frames were used as a training set, ∼170 frames were used as a validation set, and ∼140 frames were used as a test set. The mode of the yolov8 algorithm was set to “detection”. The model predictions of each individual behavioural class were validated individually on the basis of precision, recall and mean average precision with a union threshold of 0.5 (mAP50).

2.5. Data Analysis

2.5.1. Definition and Calculation of Variables

Population sizes ranged from one to sixty individuals. It was not possible to record all individuals of the larger populations simultaneously since the resolution of the frames would become too small for behaviour prediction, and the following two variables were therefore calculated from predictions of the behaviour of individuals that were located within the frames. This was accommodated by occasionally moving the frame between different parts of the population. Two variables were produced from the model’s predictions of behaviour: “proportion of time spent eating” and “proportion of time spent lying”. These were calculated as a ratio between the predicted total time spent “Eating” and “Lying” within the frames containing some or all of the population and the total time that any behaviour was predicted within the frames containing some or all of the population. This is expressed as the following two equations:

Proportion of time spent eating = Total time eating predicted/Total time behaviour predicted

(1)

Proportion of time spent lying = Total time lying predicted/Total time behaviour predicted

(2)

These variables were log- and arcsin-square-root-transformed since they are ratios, and both the non-transformed and transformed ratios were used for further statistical analysis. The number of individuals observed in each population was counted manually from thermal camera video footage, and was also used in statistical analysis.

2.5.2. Assignment of Vegetation Type to Populations of Red Deer

Each population was, by their coordinates, assigned a vegetation type from the Danish Environmental Agency or a crop type determined by a local forester. The populations were distributed in various vegetation types, including “Grey dune” (2130), “Decalcified fixed dune” (2140), “Lakeshore with small herbs” (3130), and “Humid dune slacks” (2190) as well as crop types including “Grass field”, “Corn field”, “Grain field”, “Fallow”, and “Unknown field”. However, several of the vegetation types had too few observations of red deer and were consequently removed before statistical analysis. Therefore, the final dataset comprised the following vegetation types: “Grey dune”, “Decalcified fixed dune”, “Grass field”, “Corn field”, “Grain field”, and “Unknown field”. The vegetation of “Grey dune” is dominated by herbs, grasses and lichens and is characterised by species of herbaceous plants such as Cerastium fontanum and Polygala vulgaris and species of grass such as Aira praecox and species of sedges such as Carex arenaria [43,44,45]. The vegetation type, “Decalcified fixed dune”, is a dwarf-shrub-dominated heathland and is characterised by species such as the sedge, Carex arenaria, the dwarf-shrub, Empetrum nigrum, the dwarf-shrub, Calluna vulgaris the subshrub, Genista tinctoria, and the herb, Pyrola rotundifolia [43,44,45]. Observations within the two vegetation types, “Grey dune” and “Decalcified fixed dune”, were combined under the term “natural vegetation types” in the analysis of behavioural instability. This was performed to augment the dataset by increasing the number of observations. Vegetation types were assigned to each population by their coordinates with QGIS (QGIS.org, 2024. QGIS Geographic Information System. QGIS Association. http://www.qgis.org (accessed on 5 May 2024)).

2.5.3. Statistical Analysis

All the data were imported and analysed in the statistical computing program R (“The R Project for Statistical Computing”, v. 4.3.2, https://www.r-project.org (accessed on 5 May 2024)) and in Excel. Observations more than 1.5 × IQR above the third quartile were considered outliers and consequently removed prior to any transformations and statistical analysis. To account for multiple comparisons (15 in this study), p-values were adjusted with Bonferroni correction, resulting in a significance level of p < 0.0033. All variables were tested for normality with the Shapiro test and homogeneity in variance with Bartlett’s test, and variables with at least one vegetation type lacking normality or homogeneity in variance were analysed with non-parametric tests. Only the log transformed proportion of time spent on eating showed normality of distribution in all vegetation types and homogeneity in variance between all vegetation types with sufficient sample size and were subsequently analysed with ANOVA and Tukey’s test. The non-transformed proportion of time spent eating was tested for differences between vegetation types with the Kruskal–Wallis test and subsequently with the Mann–Whitney U test. Correlation between log-transformed, arcsin-square-root-transformed as well as non-transformed proportion of time spent eating and a total number of individuals was tested with Spearman’s Rank Correlation test. MAD, kurtosis and skewness were estimated in two groups: “Natural vegetation types” and “Fields”. Differences in MAD, kurtosis, and skewness between these groups were tested by bootstrapping (n = 1000), followed by an ANOVA and Tukey’s test.

3. Results

3.1. Model Performance

The behaviours of “Eating” and “Lying” both exhibited mAP50 (mean average precision at union threshold 0.50, where a higher value indicates higher precision) values exceeding 0.7. These behaviours were the most correctly predicted behaviours, as visualised in Figure 2a. Precision, recall, mAP50, and mAP50-95 were still increasing after 30 epochs, as seen in Figure 2b, indicating that the model could have benefited from being trained for more epochs, as it would have resulted in increased performance of the model. However, since the relevant behaviour, “Eating”, as well as the behaviour “Lying”, had average mAP50 values exceeding 0.7 (see

Table 2. Summary of YOLOv8 Model Performance within each class when detecting these classes within the validation dataset consisting of 169 images. The number of instances that each class was detected within the 169 images, as well as mAP50 (mean average precision with union threshold 0.5) and mAP50-95 (mean average precision with union threshold from 0.5 to 0.95), is listed for each behaviour.

Class	Images	Instances	mAP50	mAP50-95
All	169	1042	0.704	0.448
Eating	169	231	0.85	0.67
Heatspot	169	104	0.434	0.169
Horse	169	12	0.948	0.475
Locomotion	169	62	0.502	0.315
Lying	169	289	0.77	0.421
Multiple-together	169	55	0.671	0.502
Other	169	120	0.567	0.353
Sheep	169	71	0.968	0.715
Standing	169	98	0.626	0.416

) after the 30th epoch, training was cancelled hereafter. Both the precision-confidence curves and recall-confidence curves of the behaviours “Eating” and “Lying” indicate that the model predicts these behaviours with better precision and recall than most other behaviours, as seen in Figure 2c,d. Both high precision and recall are important characteristics of the model’s predictions of the behaviour “Eating” since both falsely identifying another behaviour as “Eating” and falsely identifying the behaviour “Eating” as another behaviour will have implications for the validity of the data, and this behaviour is most crucial for examination of forage behaviour.

Since the precision and recall of the model’s predictions of the behaviours “Eating” and “Lying” were most crucial for the examination of forage behaviour, the model was accepted for use in the estimation of the proportion of time spent eating and lying.

3.2. Deer Counts in Lyngby Hede

A total of 362 deer were observed, with the most common species being red deer comprising 82.0% (297 individuals), followed by fallow deer comprising 14.4% (52 individuals), and then unknown deer species comprising 3.6% (13 individuals). Time of observation, vegetation types, deer species, as well as size and location of populations, are visualised in Figure 3. A closeup of the map covering the areas containing fields at which deer were observed is visualised in Figure 4. In total, 293 deer were observed on various crop types, with 82.9% being red deer (243 individuals), 15.7% being fallow deer (46 individuals), and 1.4% being unknown deer species (4 individuals) (see Figure A2a). Thirty-four deer were observed on “Grey dune”, with 73.5% being red deer (twenty-five individuals), 17.7% being fallow deer (six individuals), and 8.8% being unknown deer species (3 individuals) (see Figure A2b). Lastly, twenty-seven deer were observed on “Decalcified fixed dune”, with 77.8% being red deer (21 individuals) and 22.2% being unknown deer species (six individuals) (see Figure A2c).

The size of observed populations of red deer in various vegetation types can be seen in Figure 5. The vegetation type “Grey dune” encompassed a total of 25 individuals distributed across 12 populations, whereas the vegetation type “Decalcified fixed dune” encompassed a total of 21 individuals distributed across 13 populations. Deer were observed in multiple fields having four different crop types, with one of them being “Unknown fields”, encompassing 38 individuals distributed across 21 populations, “Grass field”, the crop type with the highest frequency of observations as well as the largest populations, with a total of 148 individuals distributed across seven populations, “Corn field” encompassed a total of 32 individuals distributed across only two populations, and lastly “Grain field” encompassed a total of 22 individuals distributed across two populations (see Figure 5).

3.3. Proportion of Time Spent Eating

The proportion of time spent on the behaviour “Eating” in the different vegetation types can be seen in Figure 6a. The transformed proportions of time spent eating in different vegetation types can be seen in Figure 6b,c. Generally, a higher proportion of time spent eating was observed on fields compared to natural vegetation types. The vegetation type “Grey dune” had the lowest median, and the crop type “Grain field” had the highest median in both non- and log-transformed proportions of time spent “Eating”. There was no significant difference in the proportions of time spent “Eating” between the two vegetation types, “Grey dune” and “Decalcified fixed dune”, in either the non-transformed or the two transformed proportions (p > 0.05, Mann–Whitney U test) (see Figure 6). Non-transformed proportions of time spent “Eating” were significantly lower in “Decalcified fixed dune” compared to all crop types except “Unknown fields” (p < 0.05 for all comparisons, Mann–Whitney U test) (see Figure 6a). The proportion of time spent “Eating” was significantly lower in “Grey dune” only when compared to “Grain field” (p < 0.05, Mann–Whitney U test) (see Figure 6a). In Figure 6b, the results from log-transformed proportions of time spent “Eating” were significantly lower in “Unknown fields” compared to “Grain fields” (p < 0.05, Mann–Whitney U test). The arcsin-square-root-transformed proportion of time spent “Eating” showed the same significant results as the non-transformed comparisons, with the exception of an additional significant difference between “Grey dune” and “Grass fields” (p < 0.05, Mann–Whitney U test) (see Figure 6c). None of the comparisons shown in the boxplots were significant after Bonferroni correction (p < 0.0033) (see

Table A1. Differences in the proportion of time spent on eating between combinations of vegetation types and fields tested with the Wilcoxon Ranked Sum test. Both transformations have been tested as well as the proportions themselves.

Transformation	Comparison	p-Value
Non-transformed	Grey dune—Decalcified fixed dune	p > 0.05
	Grey dune—Unknown fields	p > 0.05
	Grey dune—Grass field	p > 0.05
	Grey dune—Corn field	p > 0.05
	Grey dune—Grain field	p < 0.05
	Decalcified fixed dune—Unknown fields	p > 0.05
	Decalcified fixed dune—Grass field	p < 0.05
	Decalcified fixed dune—Corn field	p < 0.05
	Decalcified fixed dune—Grain field	p < 0.05
	Unknown field—Grass field	p < 0.05
	Unknown field—Corn field	p < 0.05
	Unknown field—Grain field	p < 0.05
	Grass field—Corn field	p > 0.05
	Grass field—Grain field	p > 0.05
	Corn field—Grain field	p > 0.05
Log-transformed	Grey dune—Decalcified fixed dune	p > 0.05
	Grey dune—Unknown field	p > 0.05
	Grey dune—Grass field	p > 0.05
	Grey dune—Corn field	p > 0.05
	Grey dune—Grain field	p > 0.05
	Decalcified fixed dune—Unknown field	p > 0.05
	Decalcified fixed dune—Grass field	p > 0.05
	Decalcified fixed dune—Corn field	p > 0.05
	Decalcified fixed dune—Grain field	p > 0.05
	Unknown field—Grass field	p > 0.05
	Unknown field—Corn field	p > 0.05
	Unknown field—Grain field	p < 0.05
	Grass field—Corn field	p > 0.05
	Grass field—Grain field	p > 0.05
	Corn field—Grain field	p > 0.05
Arcsin-square-root-transformed	Grey dune—Decalcified fixed dune	p > 0.05
	Grey dune—Unknown field	p > 0.05
	Grey dune—Grass field	p < 0.05
	Grey dune—Corn field	p < 0.05
	Grey dune—Grain field	p > 0.05
	Decalcified fixed dune—Unknown field	p > 0.05
	Decalcified fixed dune—Grass field	p < 0.05
	Decalcified fixed dune—Corn field	p < 0.05
	Decalcified fixed dune—Grain field	p < 0.05
	Unknown field—Grass field	p > 0.05
	Unknown field—Corn field	p < 0.05
	Unknown field—Grain field	p < 0.05
	Grass field—Corn field	p > 0.05
	Grass field—Grain field	p > 0.05
	Corn field—Grain field	p > 0.05

and

Table A2. Differences in the log transformation of the proportion of time spent on eating between vegetation types and field types tested with the Tukey test.

Comparison	p-Value
Grey dune—Decalcified fixed dune	p > 0.05
Grey dune—Unknown field	p > 0.05
Grey dune—Grass field	p > 0.05
Grey dune—Corn field	p > 0.05
Grey dune—Grain field	p > 0.05
Decalcified fixed dune—Unknown field	p > 0.05
Decalcified fixed dune—Grass field	p > 0.05
Decalcified fixed dune—Corn field	p > 0.05
Decalcified fixed dune—Grain field	p < 0.01
Unknown field—Grass field	p > 0.05
Unknown field—Corn field	p > 0.05
Unknown field—Grain field	p > 0.05
Grass field—Corn field	p > 0.05
Grass field—Grain field	p > 0.05
Corn field—Grain field	p > 0.05

).

The proportion of time spent on the behaviour “Eating” and the number of individuals in a population were positively correlated (ρ = 0.458, p < 0.05), indicating that the individuals tend to consume more food when they are in larger populations (see Figure 7) [46].

3.4. Proportion of Time Spent Lying

The proportion of time spent “Lying” was not significantly different between any two vegetation types (p > 0.05), indicating that the amount of rest is not dependent on the presence or quality of food, at least not under the circumstances and food availability present in the study area at the time of data collection. This was the case for both log- and arcsin-square-root transformations as well as the non-transformed proportion of time spent “Lying”.

3.5. Distributional Characteristics of Behaviours

For the analysis of behavioural instability, vegetation types were combined accordingly. “Grey dune” and “Decalcified fixed dune” were combined and labelled “Natural vegetation types”. All four crop types were combined and labelled “Fields”. All skewness values were positive, “Natural vegetation types” (1.32) and “Fields” (1.11), indicating the distribution of the proportion of time spent eating was positively skewed for both habitat categories. The kurtosis values for the proportion of time spent eating were both <3, “Natural vegetation types” (0.29) and “Fields” (−0.18), indicating platykurtic distributions for both habitat categories. The MAD of time spent eating was higher on “Fields” (0.16) compared to “Natural vegetation types” (0.01), indicating that the foraging pattern of populations of red deer was more varied on “Fields” than on “Natural vegetation types”. However, neither skewness, kurtosis, nor MAD of the proportion of time spent eating showed significant results after bootstrapping (n = 1000) and tested with ANOVA (p > 0.05). All values of skewness, kurtosis and MAD for proportion of time spent eating can be seen in

Table 3. Skewness, kurtosis and MAD (mean absolute deviation) for distributions of time spent eating in “Natural vegetation types” and “Fields”. No significant differences in skewness, kurtosis and MAD of distributions of the proportion of time spent eating were observed between the two habitat categories after bootstrapping (n = 1000) as tested with ANOVA (p > 0.05).

	Natural Vegetation Types	Fields
Skewness	1.32	1.11
Kurtosis	0.29	−0.18
MAD	0.01	0.16

.

For the proportion of time spent on the behaviour “Lying”, both skewness values were positive, “Natural vegetation types” (0.14) and “Fields” (0.49), indicating that the distribution of the proportion of time spent lying was positively skewed for both habitat categories. However, the proportion of time spent on “Lying” was more positively skewed on “Fields” compared to “Natural vegetation types”, indicating that populations located on fields had a higher variation in resting behaviour and that this variation is caused by populations resting relatively longer. The kurtosis values for time spent on the behaviour “Lying” were both <3, “Natural vegetation types” (−1.86) and “Fields” (−1.39), indicating platykurtic distributions for both habitat categories. A smaller kurtosis value for “Natural vegetation types” indicates that the time spent on the behaviour “Lying” in “Natural vegetation types” is even more spread out and flatter compared to “Fields”. This suggests that in “Natural vegetation types”, the proportion of time spent “Lying” has fewer extreme values and is more concentrated around the mean compared to “Fields”. The MAD of the proportion of time spent lying was higher on “Natural vegetation types” (0.55) compared to “Fields” (0.25), indicating that the variation of resting in populations of red deer was lower in “Fields”. However, neither skewness, kurtosis, nor MAD for the behaviour “Lying” showed significant results after bootstrapping (n = 1000) and tested with ANOVA (p > 0.05). All values of skewness, kurtosis and MAD for proportion of time spent lying can be seen in

Table 4. Skewness, kurtosis and MAD (mean absolute deviation) for distributions of time spent lying in “Natural vegetation types” and “Fields”. No significant differences in skewness, kurtosis and MAD of distributions of proportion of time lying eating were observed between the two habitat categories after bootstrapping (n = 1000) as tested with ANOVA (p > 0.05).

	Natural Vegetation Types	Fields
Skewness	0.14	0.49
Kurtosis	−1.85	−1.39
MAD	0.55	0.25

.

3.6. Time of Observation of Populations

Mann-Whitney tests showed that there was a significant difference in observation time in minutes after 7 p.m. between “Grass field” and “Grey dune” (p < 0.01), “Decalcified fixed dune” (p < 0.01 *), as well as “Unknown fields” (p < 0.01 *) as seen in

Table A3. Differences in time of observation in minutes after 7:00 p.m. between vegetation and field types.

Comparison	p-Value
Grey dune—Decalcified fixed dune	p > 0.05
Grey dune—Unknown field	p > 0.05
Grey dune—Grass field	p < 0.01
Grey dune—Corn field	p > 0.05
Grey dune—Grain field	p > 0.05
Decalcified fixed dune—Unknown field	p > 0.05
Decalcified fixed dune—Grass field	p < 0.01
Decalcified fixed dune—Corn field	p > 0.05
Decalcified fixed dune—Grain field	p > 0.05
Unknown field—Grass field	p < 0.01
Unknown field—Corn field	p > 0.05
Unknown field—Grain field	p > 0.05
Grass field—Corn field	p > 0.05
Grass field—Grain field	p > 0.05
Corn field—Grain field	p > 0.05

(“*” indicates significant difference after Bonferroni correction (p < 0.0033)). This indicates that populations were observed significantly later on “Grass fields” than on the aforementioned vegetation types. However, since the randomisation of the observation time of areas within the study area was not properly performed, the observation time is inherently biased.

4. Discussion

4.1. Methodology and Limitations Using Drone Monitoring

Camera angles directly above the observed animals were not always possible, and therefore, accurate body measurements, such as length and width, could not be obtained [1]. Consequently, species identification by object detection modelling was not performed as the data acquired were not sufficient. However, walking patterns and morphological features were discernible in the majority of the footage, facilitating manual species differentiation using thermal videos. Standardised and consistent camera angles would be optimal when recording the individuals to obtain more precise data, resulting in a less biased identification of deer species [1]. Other studies have also demonstrated the significant potential of drones for accurately identifying and monitoring animal behaviour [47]. A study by Duporge et al. 2021 also highlighted potential risks, such as disturbance caused by the drone altering the animals’ behavioural responses. In our study, sudden changes in behaviour were closely monitored when the drone flew near the animals. These potential disturbances and behavioural changes were accounted for both during the data collection and analysis. Additionally, previous studies have suggested that flying at higher altitudes reduces the risk of disturbance, a strategy that was implemented in our study [47,48]. Moreover, this study builds upon the work of Povlsen et al. (2023) and Larsen et al. (2023), who used drones to identify wild populations of species, such as deer. While their research focused primarily on species identification and the use of drones in monitoring, we expanded on their methods by also analysing behaviour and vegetation type preferences. This approach provides insights into the use of habitats, enabling differentiation between whether animals forage, rest, or migrate in the concerned vegetation type, which is essential knowledge for conservation and management of deer species [1,18]. Identification of species and behaviour requires two different camera angles. Species identification requires accurate body measures recorded from directly above, whereas behaviour identification requires footage of limbs recorded from the side. In this study, it was preferred to minimise the disturbance of the populations and to avoid observing the same individual multiple times. Therefore, flight time near deer was reduced by primarily recording the populations from the side. Moreover, species identification via modelling would have been more accessible had the study been conducted at a period when all stags were equipped with antlers, enabling differentiation based on the different shapes and sizes of the antlers between red and fallow deer. Additionally, flying during the day and identifying species with an RGB camera and switching to infrared for model training would ensure accurate species identification. Weather and seasonal variations, such as precipitation and fog, likewise impact the quality of the footage, making it more challenging to use model-based species identification. The conclusions of this study are limited by the fact that data were only collected in early spring; hence, conclusions about the annual variation in foraging and movement patterns of red deer could not be made. The thermal growth season (average daily mean temperature > 5 °C) began around the time that the data were collected [49]. Therefore, the natural vegetation types were possibly low in nutritional value, and the observed behaviour was not representative of the actual annual behavioural pattern. Winter food shortage would be expected to affect the behaviour of red deer around the time when this study was conducted. Due to logistical constraints, it was not possible to collect data across multiple seasons or account for varying vegetation growth conditions. As a result, certain seasonal behaviours and changes in vegetation were not captured in this study and should be implemented in future studies. Flying in multiple seasons and during the day would enable a comprehensive assessment of the annual and daily variation in forage behaviour. Additionally, this would illuminate whether specific growth stages of crops are particularly exposed to damage. The season also influences the detection of individuals, particularly when individuals are staying beneath tree canopies and vegetation cover [1,9]. This is presumably a larger challenge during spring (the period where this study was conducted), summer, and autumn than in the wintertime when trees are devoid of leaves.

Using drone monitoring to count individuals and populations in an area comes with the risk of observing the same individual more than once, resulting in an overestimation of population sizes [1]. An overestimation of deer populations can cause misallocation of resources and incorrect species management strategies. To minimise the risk of observing the same individual more than once in this study, takeoff spots were largely dispersed. Additionally, when observing a population, an estimation of the number of individuals was conducted, which made it possible to recognise the population in flight if observed again. It is possible that the individuals observed in the vegetation types could be the same ones observed in the fields on flights later during the night when the populations migrate to the open fields. This could be assessed by flying multiple drones simultaneously, covering both the origin and destination areas of deer migration. Individuals of red deer are known to aggregate in open areas and disperse into family groups once again when returning to forested areas [40]. The vegetation areas and crop field areas were separated by forest, possibly affecting the grouping of individuals in the migration from vegetation to field and vice versa. Populations were observed significantly later on “Grass fields” than both natural vegetation types. Herds were smaller just after sunset than later during the night (see Figure 3). Some of the herds observed on the natural vegetation types in the early night probably joined other herds later, as large populations were observed in the fields later during the night. No flights were conducted between 03:00–05:00 a.m. due to the breaks needed for recharging the drone’s batteries. Consequently, this gap in observations may lead to an incorrect perception of migration timing, suggesting that migration occurs late at night or early morning when it may actually occur earlier during the night when human activities decrease.

The amount of data collected for this study was relatively small and was only sufficient for the detection of two behaviours, “Eating” and “Lying”. Therefore, the model has been trained on a relatively small amount of data, allowing for the possibility that precision and recall would increase with a better model and making analysis of behaviour differentiated in more than just two behaviours possible. The sample size of “Corn fields” and “Grain fields” was only n = 2, likely explaining the lack of significant differences when considering the Bonferroni correction of 15 comparisons. Tracking the behaviour of individuals instead of the behaviour of populations would have increased the sample size significantly and made the determination of differences in total observed red deer on vegetation types possible.

4.2. Foraging Behavioural Patterns in Vegetation Types

The two natural vegetation types, “Grey dune” and “Decalcified fixed dune”, did not differ significantly in proportion of time spent eating, indicating that the nutritional quality of these two vegetation types may be similar to red deer. Since both vegetation types had the lowest medians (median = 0.0 for both vegetation types and n = 9 for “Grey dune” and n = 11 for “Decalcified fixed dune”), they seem not preferred as forage areas by red deer. Red deer instead foraged in nearby fields since the non-transformed proportion of time spent eating was significantly higher on multiple field types compared to the two natural vegetation types. This observation persisted in the arcsin-square-root transformations. This indicates that the nutritional status of the natural vegetation types was lower than that of the nearby fields. This potentially supports the findings of the report by the Danish Ministry of Environment and Energy, that crop damage is related to the nutritional benefits of the fields [32]. Though not significant, populations of red deer were the largest in the fields, though not significant, but were observed significantly later in “Grass field” than in the other vegetation types. These results indicate that the animals migrated to the fields to forage during the night. However, biases at the time of observation were present since the time that the different areas were covered was not randomised. Implementing a randomised schedule of flight routes would likely decrease the time bias effectively. Both the vegetation-covered areas and the fields were in close proximity to small and fragmented forests in the areas allowing for movement between the areas. This minimises the likelihood that the difference in foraging behaviour exhibited by red deer between vegetation types is due to the difference in distance to forest areas [37]. Contrarily, it suggests that the variations in foraging behaviour are more plausibly due to differences in the nutritional quality of the areas.

This study observed the highest number of individuals in “Grass field”. A study by Lande et al. (2014) supports this result, as they found that red deer prefer meadows (areas typically dominated by grass) over other vegetation types [50]. Gebert et al. (2001) described the red deer moor diet type estimated by stomach content analysis, revealing a composition primarily comprising species of Calluna, Vaccinum, and forbs [29]. The study area consisted of 70% dry and wet heath, leaving limited available space for agricultural fields in which foraging could occur. This indicates that in the absence of agricultural fields, populations of red deer in moor vegetation rely on these genera for food. A study by Jayakody et al. (2011) found that red deer have more grass included in their diet despite a much higher percentage cover of heather moor, suggesting a clear preference for grass consumption [51]. This supports the results from this study, implying that the red deer, though primarily inhabiting moors, forage on nearby grass fields or grasslands if present, after which they return to the moor. This behaviour suggests that they still prefer spending the majority of the time on the moor, as it is less disturbed than fields. These results indicate that the nutritional value and quantity of moor vegetation types are probably insufficient for sustaining populations of red deer as large as those observed in the study area, considering the red deer’s annual average dry plant matter consumption of 1.5 kg per day [31]. Therefore, mitigation of crop damage could potentially be accomplished by the establishment of grass-rich vegetation types, such as meadows, in proximity to fields. The grazing and browsing of red deer are known to contribute to the maintenance of open vegetation types [52,53]. Establishing open vegetation types near agricultural fields can mitigate crop damage caused by red deer as well as contribute to the establishment and maintenance of open natural vegetation types. The success of this method will be dependent on the nutritional value of the open natural vegetation type which, according to the results of this study, preferably should be grass rich. The amount of biomass that is removed in semi-natural grasslands and moor by red deer is comparable to that of cattle at the recommended grazing pressure [52]. It has been observed that grazing with horses and cattle simultaneously shows potential for increasing biodiversity due to differences in their dietary preferences [54]. However, a study by Scasta et al. (2016) suggests a potential diet overlap between horses, cattle, and deer, potentially impacting the biodiversity, spatial distribution, and the amount of crop damage as natural vegetation types may not provide sufficient food to sustain all populations across all species [55]. It was found that more time is spent eating with increasing population size. This would be expected, as awareness of surroundings in relation to predator avoidance is distributed across a greater number of individuals when the population is larger. Consequently, each individual can spend a greater proportion of time eating without compromising vigilance. Rowe et al. (2023) found that the blink rate of individuals correlated positively with group size, indicating reduced vigilance and increased safety with larger herds [56]. Other studies support the hypothesis that bigger herd sizes are an effective anti-predator strategy, particularly in open areas such as fields [57]. No significant difference in the proportion of time spent lying was observed between any two vegetation types. However, due to a low sample size for several crop types, no definite conclusions can be made regarding patterns in the resting behaviour of red deer depending on vegetation type. This behaviour was included in statistical analysis not because it was deemed highly relevant for the analysis of forage behaviour but rather because of the model’s high precision and recall in predictions of it.

4.3. Behavioural Instability

Skewness values for both behaviours were positive for all vegetation types, indicating a higher concentration of the data in the low values and, therefore, below the mean [58]. The results showed that the distribution of the proportion of time spent eating was more positively skewed for “Natural vegetation types” than “Fields”, indicating that the eating behaviour tends to be more concentrated towards shorter durations in “Natural vegetation types” compared to “Fields”. This suggests potential differences in foraging patterns or resource utilisation between these two habitat categories. However, the difference was not significant. Furthermore, kurtosis values were all less than 3, which is a platykurtic distribution, suggesting few and less extreme values in the two behaviours within the observed populations [58]. Conclusively, the behaviours had high variation, with most of the variation being caused by populations with a high proportion of time spent on both behaviours. Despite observing differences in skewness, kurtosis, and MAD between the two different habitats, no significant differences were found after bootstrapping and ANOVA testing. The lack of significance could be due to various factors, such as small sample sizes or unaccounted variables that may have affected behaviour.

5. Conclusions

The study of forage behaviour in red deer using a drone equipped with a thermal camera combined with object detection models has proved to be an efficient method for the identification of behaviours. While promising, this method has several challenges limiting its applicability. The method is somewhat limited to open vegetation types, especially during the growing season, due to low visibility through tree canopies. Randomisation of takeoff time and spots would decrease time bias and increase the precision of estimates of migration patterns and timing. Despite these challenges, this study found that populations of red deer spent a smaller proportion of time eating in the “Grey dune” and “Decalcified fixed dune” compared to multiple crop types. The significant difference in the time of observations between the “Grass field” and the two natural vegetation types indicates that migration occurred during the night from the naturally vegetated areas towards the fields and that the smaller family units probably were aggregated into larger groups when migrating to the fields. This indicates that the nutritional status of the relatively lean vegetation types was considered insufficient by the red deer, necessitating their migration to the fields. It also indicates that red deer prefer to reside in relatively preserved vegetation types, possibly due to less disturbance from anthropological activity. However, this study was temporally limited to early spring and nighttime, necessitating studies on the vegetation preferences in red deer in Denmark across different seasons and during the daytime.