Mapping Waterbird Habitats with UAV-Derived 2D Orthomosaic Along Belgium’s Lieve Canal

Abstract

The accurate monitoring of waterbird abundance and their habitat preferences is essential for effective ecological management and conservation planning in aquatic ecosystems. This study explores the efficacy of unmanned aerial vehicle (UAV)-based high-resolution orthomosaics for waterbird monitoring and mapping along the Lieve Canal, Belgium. We systematically classified habitats into residential, industrial, riparian tree, and herbaceous vegetation zones, examining their influence on the spatial distribution of three focal waterbird species: Eurasian coot (Fulica atra), common moorhen (Gallinula chloropus), and wild duck (Anas platyrhynchos). Herbaceous vegetation zones consistently supported the highest waterbird densities, attributed to abundant nesting substrates and minimal human disturbance. UAV-based waterbird counts correlated strongly with ground-based surveys (R² = 0.668), though species-specific detectability varied significantly due to morphological visibility and ecological behaviors. Detection accuracy was highest for coots, intermediate for ducks, and lowest for moorhens, highlighting the crucial role of image resolution ground sampling distance (GSD) in aerial monitoring. Operational challenges, including image occlusion and habitat complexity, underline the need for tailored survey protocols and advanced sensing techniques. Our findings demonstrate that UAV imagery provides a reliable and scalable method for monitoring waterbird habitats, offering critical insights for biodiversity conservation and sustainable management practices in aquatic landscapes.

1. Introduction

Aquatic ecosystems, particularly river systems, play a fundamental role in maintaining biodiversity by providing essential habitats for a wide range of aquatic and semi-aquatic species [1,2,3]. Among them, waterbirds serve as important bioindicators due to their strong dependence on aquatic habitats and their sensitivity to environmental changes [4,5]. The effective monitoring of waterbird populations is important for assessing the ecological health of river systems, detecting habitat degradation, and informing conservation strategies [6,7]. Previous studies, such as traditional aerial surveys and ground-based surveys, are among the most commonly used methods [8,9,10,11]. Although ground-based surveys relying on manual visual observation tend to yield relatively accurate results, they are labor-intensive, susceptible to observer bias, spatially constrained, and associated with high operational costs [12,13]. In addition, these methods often cause disturbance to wildlife. Many of the target bird species prefer breeding in remote and inaccessible areas, making ground surveys impractical without causing significant disturbance to breeding individuals [11,14]. Traditional aerial surveys, such as those based on satellite imagery or helicopter-based aerial photography, can serve as alternatives by enabling efficient large-scale monitoring in areas that are difficult for humans to access. However, they are limited by their relatively low spatial resolution, which generally makes them unsuitable for high-resolution ecological investigations at local scales [8,9,11]. These limitations are particularly pronounced in artificial river systems, where waterbirds utilize diverse microhabitats with varying levels of accessibility [15,16]. In such complex and heterogeneous environments, traditional survey techniques often lack the spatial and temporal resolution required for precise ecological assessments [11,17,18]. Thus, there is a pressing need to develop more efficient, scalable, and minimally invasive monitoring approaches that can improve both data accuracy and survey efficiency.

The latest advancements in Unmanned Aerial Vehicles (UAVs) have enhanced wildlife monitoring by offering high-resolution, repeatable, and cost-effective survey methods [19,20]. Compared to conventional techniques, UAVs provide broader spatial coverage, reduce observer-related biases, and minimize disturbances to bird populations. Unlike satellite remote sensing, which is constrained by low spatial resolution and atmospheric interference, UAVs operate at low altitudes, capturing fine-scale details of bird distributions and habitat conditions [11,17,18]. Previous studies have demonstrated the efficacy of UAVs in mapping wetlands, lakes, and floodplains, where dense vegetation and inaccessible nesting sites hinder ground-based observations [21,22,23,24,25]. UAV-derived orthomosaics offer precise spatial datasets that facilitate species identification, population estimation, and habitat characterization, enhancing our ability to monitor long-term ecological trends and assess the impacts of anthropogenic pressures on waterbird communities [26,27,28]. While UAV technology has been widely applied in monitoring water quality, vegetation dynamics, and habitat structure in aquatic environments, its specific application to waterbird monitoring remains relatively underexplored, particularly due to challenges such as detecting small, vegetation-concealed species and the influence of environmental conditions on image accuracy. Given the behavioral complexity, high mobility, and ecological sensitivity of waterbird species, further research is needed to evaluate the accuracy, efficiency, and suitability of UAV-based approaches for avian population assessment [17,29,30].

In addition to enhancing waterbird census methodologies, UAV-derived orthomosaics provide a powerful tool for high-resolution habitat mapping. These images allow researchers to identify critical nesting and foraging sites, quantify habitat modifications, and track ecological changes over time [11,17]. Such capabilities are especially valuable in riverine landscapes, where hydrological processes and seasonal fluctuations affect habitat availability [31,32]. Despite the clear advantages of UAV-based monitoring, the standardization of methodologies for riverine waterbird surveys remains inadequate, with flight parameters, sensor specifications, image processing workflows, and validation protocols still requiring careful optimization to ensure data accuracy and reproducibility [11,29,30]. As UAV technology continues to be recognized as a transformative tool in ecological research, the establishment of standardized protocols will be critical for integrating UAV-derived data into long-term ecological monitoring programs [27,33,34].

The Lieve canal, one of Europe’s earliest artificial canals [35] situated in Flanders, Belgium, provides a valuable location for assessing UAV-based waterbird monitoring due to its controlled hydrology, open channel structure, and diverse bird assemblages within a human-modified freshwater habitat [15,36,37]. Originally constructed in the 13th century for transport, this historical waterway has undergone significant ecological transformations, currently functioning as a highly diverse aquatic habitat that supports a variety of waterbird species [36,37]. Among the dominant species inhabiting the Lieve canal are the Eurasian coot (Fulica atra), common moorhen (Gallinula chloropus), and wild duck (Anas platyrhynchos), each exhibiting distinct habitat preferences [16]. While several species rely on dense emergent vegetation for nesting, others depend on open-water zones for foraging [38,39]. However, human activities such as agriculture, wastewater discharge, and recreation may impact habitat quality and contribute to observed patterns in waterbird distribution along the Lieve canal [15,40,41]. Given these ecological dynamics, understanding the spatial and temporal distribution of waterbirds and their habitat associations is essential for developing conservation strategies and ensuring sustainable river management.

In this study, we developed a UAV-based approach to map waterbird habitats and assess species-specific habitat preferences along the Lieve canal during the nesting season. High-resolution UAV-derived 2D orthomosaics were used to classify land cover and delineate habitat types across all survey sections. Waterbird abundance was then quantitatively analyzed to infer habitat associations. Subsequently, UAV-derived bird counts were validated against ground surveys, and detection performance was evaluated with respect to image resolution and species traits. By integrating remote sensing, habitat classification, and waterbird behavioral ecology, this approach provides a scalable framework for linking species monitoring with fine-scale habitat assessment in managed aquatic landscapes.

2. Materials and Methods

2.1. Study Area

This study was conducted along an 11 km section of the Lieve canal, located in Lievegem, near Ghent, Belgium (51°03′ N, 3°43′ E) (Figure 1). The study area encompasses diverse land-use types, including agricultural fields, suburban developments, and urbanized zones, all of which contribute to varying environmental pressures, such as nutrient runoff, wastewater discharge, and recreational activities [3,15]. These factors influence water quality and habitat conditions along the river. The region is topographically flat, with an elevation of approximately 30 m above sea level, and has a temperate maritime climate, which supports extensive aquatic ecosystems and agricultural activities [42].

To ensure comprehensive monitoring, the river was divided into 24 study sections, labeled F1 to F24, each approximately 400 m in length, except for F17, which spans 350 m. Both UAV surveys and ground-based bird counts were conducted at regularly spaced sampling points within these sections to capture spatiotemporal variations in waterbird distribution and habitat characteristics. Figure 1 provides a detailed map of the study area, illustrating the land-use patterns and the locations of sampling sites. This spatial division ensures consistent coverage across the study area, enabling a systematic analysis of local ecological dynamics and the relationships between waterbird abundance, water quality, and habitat conditions. Each sampling point indicates the start of the 400 m study section, which runs approximately up until the next sampling point.

2.2. Data Collection

2.2.1. UAV Data Collection

We used the DJI Mavic 3 Multispectral (M3M) UAV (DJI Technology Co., Ltd., Shenzhen, China), equipped with a high-resolution 4/3 CMOS sensor capable of capturing 20-megapixel (20MP) RGB images for aerial data collection and a mechanical shutter with a maximum speed of 1/2000 s minimized motion blur during image acquisition. The UAV’s Real-Time Kinematic (RTK) module provides centimeter-level positioning accuracy, ensuring the precise georeferencing of captured images. Its advanced stabilization system maintains consistent image quality even under moderate wind conditions [43]. The camera captured images in RGB format, suitable for detecting individual waterbirds and mapping habitat characteristics [44].

Flight missions were planned and executed using the DJI Pilot 2 software (Version 10.1.0.30, DJI Technology Co., Ltd., Shenzhen, China), ensuring precise and repeatable flight paths. An example of a planned flight pattern is provided in Figure S1 (Supplementary Materials). The survey design followed a grid-based approach, covering all 24 study sections (F1–F24). The standard flight altitude was set at 25 m (GSD: 0.67 cm/pixel) to balance image resolution and coverage efficiency, with site-specific adjustments based on environmental constraints. In areas with tall obstacles such as trees and power lines (e.g., F1, F2, F5, F6), the flight altitude was increased to 35 m (GSD: 0.94 cm/pixel) to prevent collisions. Conversely, in open landscapes (e.g., F9, F19, F20, F21, F22, and F24), the flight altitude was lowered to 20 m to enhance spatial resolution (GSD: 0.54 cm/pixel). In F23, due to the presence of high-voltage power lines, the UAV was operated at 15 m (GSD: 0.40 cm/pixel) for safety considerations. The GSD is inherently related to flight altitude, sensor width, focal length, and image width [45]. In this study, GSD values were automatically calculated by the DJI Pilot 2 flight control system, based on camera resolution, flight height, and other flight parameters. Detailed information on flight altitude settings per location is summarized in Table S1 (SM). The UAV followed a pre-programmed autonomous flight route, with an 80% forward overlap ratio and a 70% side overlap ratio, ensuring high-quality orthophoto reconstruction during post-processing [45]. All flights were conducted in March 2024 under optimal weather conditions to minimize disturbances from wind, rain, or low visibility. The UAV’s camera settings were automatically adjusted to dynamic lighting conditions to maintain consistent exposure across all images.

To ensure accurate georeferencing, multiple complementary techniques were employed. The UAV’s RTK module utilized Global Navigation Satellite System (GNSS) data and was integrated with the FLEPOS 3.0 real-time network correction service (Flemish Positioning Service, Government of Flanders, Belgium). This system ensured real-time positional adjustments, enhancing spatial accuracy. Additionally, Ground Control Points (GCPs) were systematically deployed to further improve orthophoto reconstruction accuracy [45,46,47]. In each flight zone, GCPs were placed every 50 m, ensuring uniform spatial coverage. Circular yellow ground markers were used as GCPs, with a black center mark, created by applying yellow and black paint for precise identification in UAV imagery. The coordinates of all GCPs were measured using an Emlid Reach RS+ handheld GPS receiver (Emlid Ltd., Saint Petersburg, Russia), using the FLEPOS NTRIP service, providing highly accurate reference points for georectification.

2.2.2. Waterbird Survey

Waterbird monitoring was conducted using two complementary methods: ground-based visual counting and UAV-based image analysis. Ground-based surveys were conducted in accordance with standardized ornithological protocols to provide a comparative dataset [48,49]. Individual waterbirds were identified based on morphological characteristics, such as body shape, size, and spatial clustering patterns. Three focal species were surveyed: Eurasian coot (Fulica atra), common moorhen (Gallinula chloropus), and wild duck (Anas platyrhynchos). For clarity, these species are hereafter referred to as coot, moorhen, and duck, respectively. The study area was systematically divided into 24 monitoring transects, each corresponding to a 400 m UAV flight zone, which was further subdivided into eight 50 m segments to enable higher-resolution ground-based bird counts. To ensure accuracy and prevent duplicate counting, each transect was clearly marked, and observers followed a structured data collection approach. Field observers conducted direct visual counts at predefined observation points, recording species and abundance following standardized data sheets, with an example of such records provided in Table S2 (SM). To minimize disturbance, surveyors maintained a consistent distance from waterbirds to avoid altering natural behavior.

2.3. Data Analysis

2.3.1. UAV Orthomosaic Generation

The UAV image reconstruction process was conducted using Agisoft Metashape Professional (Version 2.1.2, Agisoft LLC, St. Petersburg, Russia) to ensure accurate georeferencing and high spatial resolution, following established photogrammetric workflows based on Structure-from-Motion (SfM) techniques commonly applied in UAV- derived orthomosaic generation studies [50,51,52,53]. The process began with low-accuracy initial image alignment and tie point detection, generating a sparse point cloud from overlapping images. GCPs were then incorporated to improve spatial accuracy, followed by camera optimization to refine both internal and external parameters, minimizing lens distortion and reducing spatial errors. Subsequently, a second round of high-accuracy alignment and point cloud densification was performed, enhancing the consistency of image geometry and increasing the level of spatial detail.

Orthorectification and mosaic generation were conducted using internal image geometry reconstructed through the SfM process, based on tie points and optimized camera parameters. This approach eliminated the need for generating a digital elevation model (DEM), thereby significantly reducing processing time while ensuring the spatial accuracy required for ecological applications [50,54,55,56]. Given the relatively flat terrain across the study area and the primary research focus on acquiring waterbird imagery rather than modeling topographic variation, DEM-based correction was deemed unnecessary [51,55,57]. Moreover, previous findings suggest that assuming flat terrain may yield more reliable spatial outputs in open vegetated landscapes than DEM-based correction, due to noise and structural inconsistencies in 3D reconstruction [45,58,59]. The resulting UAV-derived orthomosaic provided seamless, high-resolution 2D spatial representations suitable for subsequent waterbird detection and habitat mapping.

After generating the UAV-derived orthomosaics, their geometric quality was assessed by calculating the root mean square error (RMSE) based on the residuals at the GCPs. The RMSE quantified the spatial alignment between the reconstructed positions of the GCPs in the orthomosaic and their corresponding ground-truth coordinates, as recorded during field acquisition. This assessment served as a practical indicator of geometric consistency and alignment precision, supporting the reliability of the subsequent spatial analyses. The application of GCP-based RMSE metrics remains a widely accepted approach in UAV-based remote sensing and ecological monitoring workflows [54,60,61]. Additionally, the orthomosaics were inspected in detail to identify distortions, stitching artifacts, and resolution inconsistencies that may affect the quality of the generated model, as emphasized by previous studies [62,63,64].

2.3.2. Habitat Classification

Habitat classification was conducted using high-resolution orthomosaic imagery within the ArcGIS 10.8 software (Esri, Redlands, CA, USA), a widely used platform for spatial analysis and environmental mapping [65,66,67]. Following established remote sensing protocols, an automated unsupervised ISO clustering classification was applied to orthomosaic imagery from all 24 UAV-monitored locations, followed by the GIS-based visual correction and vector delineation of land cover patches to improve accuracy. This hybrid approach, widely used in fine-scale landscape ecology, balances the efficiency of automated classification with the precision of expert interpretation in complex habitats where spectral similarity may limit algorithmic reliability [66,68,69]. Distinct land cover categories were identified based on texture, tone, and contextual features, including residential infrastructure, industrial facilities, farmlands, tree cover, herbaceous vegetation (e.g., reeds), open water bodies, and linear features such as roads. The proportional coverage of each class was calculated per section, providing quantitative inputs for ecological stratification. Based on the dominant land cover composition at each study section, all 24 monitoring sections (F1–F24) were then classified into four ecologically meaningful habitat types: (A) residential zone, (B) industrial zone, (C) riparian tree zone, and (D) herbaceous vegetation zone. Notably, although section F7 was not internally dominated by industrial features, it was nonetheless classified as industrial zone due to the continuous presence of adjacent agro-industrial land use at both its upstream and its downstream boundaries. This classification supports standardized environmental mapping to support spatial analyses of waterbird–habitat interactions. It generates georeferenced, repeatable, and auditable habitat datasets suitable for long-term assessments and cross-site comparisons.

2.3.3. Waterbird Annotation from Orthomosaics

Bird detection on UAV-derived orthomosaics was conducted through systematic manual interpretation using QGIS 3.28 (QGIS Development Team, Open Source Geospatial Foundation, Beaverton, OR, USA). To ensure consistency across locations, all orthomosaics were examined at a fixed viewing scale (approximately 1:100–1:150), corresponding to GSDs ranging from 0.40 cm/pixel to 0.94 cm/pixel, depending on flight altitude (15 m, 20 m, 25 m, or 35 m). Detection was based on clear visual cues such as body outline, contrast, and shadows distinguishable from surrounding features. All annotations were performed by a single trained observer following a consistent protocol to minimize subjectivity. To ensure reliability, approximately 20% of the annotations were independently cross-checked by a second observer. This structured approach ensured annotation consistency and provided a robust basis for comparing orthomosaic-based and ground-based survey results [70,71,72].

2.3.4. Statistical Analysis

To investigate the capability and accuracy of UAV-based counts in detecting waterbirds, we compared the results obtained from UAV imagery with those from ground-based visual surveys. Detection accuracy was first calculated individually for each flight area (F1–F24) based on UAV and ground counts. The overall detection accuracy for the survey was then determined by averaging the detection accuracy values across all flight areas within each GSD group, following approaches used in ecological studies under small-sample conditions [73,74]. The detection accuracy of UAV-based counts was quantified using Equation (1), as follows:

$$Accuracy=\min \left({\displaystyle \frac{y_{(i,\mathrm{UAV})}}{y_{(i,\mathrm{ground})}}},1\right)\times 100\%$$

(1)

where $y_{(i,\mathrm{UAV})}$ represents the number of individuals detected by UAV imagery at observation i and $y_{(i,\mathrm{ground})}$ represents the number recorded during ground-based surveys.

In practical field surveys, UAV-based counts occasionally exceeded ground-based counts. This situation is reasonable because waterbirds may move between adjacent flight areas, especially near the boundaries, resulting in an apparent increase in the UAV-based count for a given area. Since the primary objective of this study was to evaluate the UAV’s capability to detect the presence of waterbirds, such cases were interpreted as successful detections without penalization. Therefore, when the UAV count exceeded the ground count for a specific flight area, the detection accuracy was still assigned a value of 100%. This approach ensured that the assessment focused on the UAV’s ability to identify all individuals recorded by ground-based observations while allowing for minor variations caused by the natural movement of birds during the survey. A special case also needs to be considered: when both the UAV-based and ground-based counts are zero, it indicates that no waterbirds were detected in the surveyed area by either method. In this situation, the UAV count is consistent with the ground-based observation, and the detection accuracy is therefore assigned as 100%.

To evaluate differences in waterbird abundance among the four habitat types (A = residential zone, B = industrial zone, C = riparian tree zone, and D = herbaceous vegetation zone), we used non-parametric Kruskal–Wallis tests given the limited sample size (n = 5–7 per group) and violations of normality in some groups. When significant differences were detected, pairwise comparisons were performed using Dunn’s post-hoc test with Bonferroni correction to identify which habitat categories differed significantly. This approach offers a robust alternative to ANOVA for small and non-normally distributed ecological datasets. All analyses were conducted using R version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria) via RStudio (version 2022.12.0, Posit Software, Boston, MA, USA), with statistical significance set at α = 0.05.

In addition, linear regression analysis was conducted to evaluate the agreement between UAV-based and ground-based bird counts across all monitored locations. Separate models were built for each species (coot, duck, and moorhen), as well as for the total waterbird count. Ground count values were used as the independent variable, and UAV-derived counts as the dependent variable. Model performance was evaluated using the coefficient of determination (R²) to quantify the proportion of variance explained, and 95% confidence intervals were visualized to indicate the reliability of each regression fit.

3. Results and Discussion

3.1. High-Resolution Habitat Classification Using UAV Orthomosaics

Figure 2A–D illustrate the four main habitat types classified along the Lieve Canal, with representative examples extracted from high-resolution UAV orthomosaics: Residential zones, Industrial zones, Riparian tree zones, and Herbaceous vegetation zones. Detailed classification results are provided in Table S4 (Supplementary Materials). Residential zones were characterized by dense housing, hardened surfaces, simplified vegetation structure, and frequent human presence. Industrial zones were located near factory and storage facilities, often associated with visible anthropogenic disturbance and potential sources of pollution. Riparian tree zones consist of open areas without built infrastructure, where mature trees, such as European beech (Fagus sylvatica), plane trees (Platanus spp.), and English oak (Quercus robur), form continuous and structurally complex riparian belts. Herbaceous vegetation zones were composed primarily of grasses (Poaceae), sedges (Cyperaceae), reeds (Phragmites spp.), and low shrubs, with no apparent canopy trees or built structures, and were generally situated in open terrain.

Quantitative analysis of vectorized land cover maps reveals clear structural differences among habitat types (Figure 3). In the residential zone (e.g., F17), residential buildings and roads jointly accounted for 36.7% of the area, indicating a high degree of urbanization with frequent human activity, noise, and domestic animal presence, all of which may influence waterbird behavior and habitat selection. The industrial zone (e.g., F6) exhibited a combination of industrial infrastructure and paved surfaces totaling approximately 19.9%, suggesting occasional heavy machinery operation and restricted public access, which may reduce direct human disturbance but contribute to broader environmental impacts such as runoff or noise. The riparian tree zone (e.g., F16) featured a notable 11.6% tree cover, offering potential shelter and vertical structure for some species. In contrast, the herbaceous vegetation zone (e.g., F20) was dominated by low vegetation such as grass and reeds (14.2%), with no tree canopy or built structures, providing an open and minimally disturbed foraging environment. These habitat-specific variations in land cover composition reflect distinct levels of human disturbance and vegetation structure, offering critical environmental context for interpreting spatial patterns of waterbird occurrence across the study area.

In addition, the study provides a comprehensive analysis of vectorized land cover maps for all 24 surveyed sections (Figures S-F1–S-F24 in the SM). Across the full study area, farmland emerged as the dominant background habitat, occupying an average of 56.8% of the total area, followed by grassland (13.7%), waterbodies (10.8%), tree cover (6.8%), and industrial areas (1.5%) (Figure S2, SM). A further classification of the sites into four habitat types revealed distinct landscape characteristics: in residential zones, areas of human activity, including residential houses and paved surfaces, accounted for 25.8%; in industrial zones, built-up and impervious surfaces represented 11.2%; in riparian tree zones, tree cover exceeded 10%, making it the most prominent feature aside from farmland; and in herbaceous vegetation zones, grassland reached 23.2% coverage (Table S5, SM). The resulting habitat classification maps provide an intuitive visual representation of habitat structure and inter-site variability, offering a valuable basis for interpreting waterbird nesting behavior and habitat selection. Furthermore, these maps serve as a spatially explicit reference for local water management authorities along the Lieve Canal, contributing to long-term monitoring and ecological decision making.

The residual deviations at the GCPs used in model construction yielded RMSE values of 1.34 cm (X), 1.55 cm (Y), and 0.71 cm (Z), with an aggregated three-dimensional error of 2.35 cm. The mean image reprojection error across the bundle adjustment was 0.83 pixels (Supplementary Material, Table S3). The observed results in this study aligns with previous UAV-based ecological monitoring applications, where GCP-based RMSE values typically range between 1–3 cm in the horizontal plane and 1–5 cm in the vertical plane [46,75,76]. The findings demonstrate a high degree of geometric consistency, which supports the reliability of the image alignment for subsequent ecological interpretation. The stable image reprojection error (0.83 pixels) further indicates consistent camera calibration and tie-point matching quality throughout the flights. The observed geometric consistency is attributed to the systematic GCP distribution and standardized flight parameters, including a forward overlap ratio of 70% and a side overlap ratio of 80%, which enhanced image redundancy and alignment [46,47,77].

However, these values do not constitute an independent assessment of absolute geolocation accuracy. Due to practical constraints in the field, independent ground checkpoints were not established in this study. Although some studies have suggested that they may not always be essential [78,79,80], future research in similar ecological settings would benefit from incorporating independent checkpoints to enable more refined accuracy assessments [45,81,82].

3.2. Waterbird Abundance Patterns Across Habitat Types

The spatial distribution of waterbirds along the Lieve Canal exhibited marked heterogeneity across survey sections. As shown in Figure 4, the highest local abundance was recorded at site F9, with an average density of 40 individuals per kilometer, significantly exceeding that of other segments. According to the UAV-based habitat classification, F9 represents a typical example of the D-type herbaceous vegetation zone, characterized by open space devoid of residential or industrial structures and dominated by low-stature vegetation such as grasses and reeds. Similarly, the downstream segment encompassing sites F19 to F23, also classified as herbaceous vegetation zones, showed consistently high and stable densities, averaging 26 individuals per kilometer. All three monitored waterbird species were present in this segment, suggesting structurally rich and functionally suitable conditions during the breeding season. The continuous presence of herbaceous vegetation along the river margins in these zones likely offers ample nesting opportunities, aligning with findings from previous studies. In contrast, low waterbird densities were observed in residential areas such as F1, F3, and F8, where frequent human activity, dense housing, and paved surfaces may act as disturbance factors limiting bird occurrence.

Further statistical analysis (Figure 5) confirmed that the herbaceous vegetation zones (D) supported the highest waterbird densities, with significant differences compared to the other habitat types. Although no statistically significant differences were found among the residential zones (A), industrial zones (B), and riparian tree zones (C), their average densities varied notably. The lowest densities were observed in the C-type riparian tree zones, which are characterized by uniformly distributed tall trees. In these areas, the growth of understory vegetation such as grasses and reeds may be suppressed due to shading, limiting the availability of suitable nesting substrates. Residential (A) and industrial (B) zones showed comparable densities, both subject to more intense anthropogenic disturbance compared to D zones. However, unlike the riparian tree zones, several A and B sections, including F1, F7, and F18, were observed to host submerged macrophytes based on high-resolution UAV orthomosaics. These aquatic plant resources are likely important for species such as the Eurasian coot, which rely on submerged vegetation for foraging. Overall, the observed spatial distribution patterns of waterbirds reflect the influence of habitat type on waterbird occurrence and behavior along the canal corridor.

3.3. Validation of UAV-Based Waterbird Counts

3.3.1. Species-Specific Counting Performance

The comparison between UAV-derived orthomosaic counts and ground-based observations revealed distinct interspecific differences in detection accuracy (Figure 6). Among the three target species, the Eurasian coot showed the highest concordance between aerial and ground-based counts, with a coefficient of determination (R²) of 0.801. Wild duck followed with R² = 0.718, also indicating a strong correlation. In contrast, common moorhen counts exhibited the weakest correspondence, with R² = 0.475, falling below the conventional threshold of 0.5. When aggregating all waterbird counts across monitoring sections, the overall correlation remained moderately strong (R² = 0.668), suggesting that UAV-based surveys offer a reliable alternative to traditional ground-based methods for estimating total bird abundance.

The observed variation in detection performance among species likely stems from differences in behavioral ecology, flocking tendency, and visual detectability. As detailed in

Table 1. Summary of body size characteristics and distinctive features of coot (Fulica atra), moorhen (Gallinula chloropus), and duck (Anas platyrhynchos).

Species	Feature	Feature Description	Reference
Coot	White frontal shield	Width: 1.5–3.1 cm	[85,86,87]
Moorhen	Red frontal shield	Width: 1.0–1.5 cm in adults; 0.4–1.0 cm in juveniles	[84,88]
Duck	Dark, iridescent-green head; white-bordered blue speculum	Head width: 3.5 cm; Speculum occupying: 1/3–1/2 of folded wing surface	[89,90,91,92]

, coots are characterized by a prominent white frontal shield (1.5–3.1 cm wide), which enhances their visibility in aerial imagery. In contrast, moorhens possess a much smaller frontal shield (1–1.5 cm in adults, as small as 0.4 cm in juveniles), rendering them more cryptic, particularly under suboptimal lighting or vegetation cover. Although ducks are larger and exhibit more distinctive morphological features, their nesting behavior often leads them away from open water into bank-side vegetation during the breeding season, reducing their detectability in UAV orthomosaics. These findings are consistent with previous studies [6,17,33,83,84], which demonstrate that wildlife bird surveys and modeling should comprehensively consider the external features and body structure of migratory birds, along with realistic body size scaling. Moreover, species-specific traits such as interspecific morphological differences and behavioral variation need to be fully considered. These findings underscore the importance of considering species-specific traits when applying remote sensing approaches to waterbird monitoring.

3.3.2. Effects of GSD on UAV-Based Counting Performance

Detection performance under different image resolutions was evaluated by comparing UAV counts across three GSDs: 0.54 cm, 0.67 cm, and 0.94 cm (

Table 2. Accuracy of bird identification based on UAV counts at different GSD.

GSD (cm/pixel)	Coot (%)	Moorhen (%)	Duck (%)	Total (%)
0.54	100.00	80.56	28.33	80.56
0.67	79.17	38.97	82.69	66.16
0.94	60.00	16.67	100.00	37.50

). Overall detection accuracy declined markedly as GSD increased, dropping from 80.56% at 0.54 cm to 37.50% at 0.94 cm, underscoring the critical role of image resolution in accurate species identification. Both coots (Fulica atra) and moorhens (Gallinula chloropus) exhibited a pronounced decrease in detection accuracy with lower image resolution. For coots, accuracy declined from 100.00% to 79.17% and further to 60.00%. Moorhen detection was more severely impacted, with accuracy falling from 80.56% to 38.97% and then to 16.67%. In contrast, ducks (Anas platyrhynchos) exhibited a different trend, with detection accuracy remaining relatively stable across GSDs and even showing a lower value at the highest resolution. Unlike coots and moorhens, ducks showed relatively stable detection accuracy at coarser GSDs (0.67–0.94 cm), likely due to their larger body size and distinct coloration. A decline at the finest resolution (0.54 cm) may reflect sample scarcity, as ducks were less abundant and often outside the imaged areas due to their preference for nesting away from open water. Potential disturbance effects cannot be excluded and merit further study. These findings highlight the species-specific sensitivity to image quality and reinforce the importance of high-resolution imagery in UAV-based waterbird monitoring.

Figure 7 presents representative image snippets of moorhens across three resolutions. At GSD = 0.54 cm, all three key visual features, red frontal shield, white flanks, and white undertail patch, are clearly distinguishable, supporting accurate identification. As GSD increases to 0.67 cm, fine features such as the frontal shield become partially obscured, leaving only the white flank and undertail patch discernible. At the coarsest resolution (GSD = 0.94 cm), only one white marking remains marginally visible, and the overall body outline becomes blurred, leading to a substantial loss of identification cues. Notably, these differences were assessed under a constant 1200% zoom level to ensure consistent visual comparison. The progressive loss of feature visibility with decreasing resolution explains the sharp drop in moorhen detection accuracy from 80.56% to 16.67% (Table 2) and underscores the critical role of GSD in determining the detectability of small-bodied or visually cryptic species in UAV orthomosaics.

4. Discussion

4.1. Ecological Insights into Waterbird Habitat Selection Based on UAV-Derived Habitat Mapping

Our findings underscore the potential of UAV-based remote sensing, particularly high-resolution orthomosaic imagery, for advancing ecological understanding and biodiversity conservation in aquatic habitats. During the nesting season, when waterbirds exhibit relatively stable site fidelity and defined habitat preferences, species tend to select breeding sites that offer a combination of reliable food accessibility, suitable nesting substrates, minimal anthropogenic disturbance, and sufficient concealment from predators. A range of habitat elements, including shoreline vegetation structure (e.g., dominance of tall trees versus low-stature reeds and grasses), the availability of submerged macrophytes, bank exposure, adjacent land use intensity, and human activity levels, exert a direct influence on the habitat selection decisions of breeding waterbirds. These factors not only shape the physical structure of available habitats but also determine ecological quality in terms of nesting success and foraging efficiency.

In this study, high-resolution orthomosaics produced from UAV imagery enabled the detailed classification of four dominant habitat types along the Lieve Canal. Among these, herbaceous vegetation zones were most consistently associated with elevated waterbird densities. These zones, typically characterized by open spatial configuration, dense graminoid cover such as Phragmites and Carex, and the absence of built structures, provided ideal conditions for nesting and foraging. Conversely, residential and industrial zones were subject to elevated levels of human disturbance, including noise and infrastructure, which appeared to reduce their suitability for breeding birds. Riparian tree zones, while vegetated, were dominated by tall canopy trees that limited the development of understory vegetation due to shading, thereby constraining nesting opportunities and food availability. Collectively, these results demonstrate that UAV-enabled habitat mapping offers a powerful tool for linking fine-scale habitat features to species distribution patterns and provides actionable insights for conservation practitioners and ecosystem managers aiming to protect waterbird biodiversity in managed canal environments.

4.2. Factors Influencing UAV Counting Performance

The ability of UAVs to detect and count waterbirds with high accuracy is influenced by a combination of technical, biological, and environmental factors. Among these, image spatial resolution, determined by GSD, flight altitude, species-specific behavioral traits (e.g., activity patterns, territoriality, and habitat preferences), and the structural complexity of the surrounding habitat all play crucial roles. Our findings confirm that GSD is a primary determinant of detection performance. Higher spatial resolution significantly enhances the clarity and boundary definition of avian targets, thereby improving species recognition and classification accuracy. Consequently, UAV flight planning should be adapted to the typical morphological characteristics of the target species (e.g., body size, plumage contrast, or posture), with flight height and camera settings calibrated to ensure that the captured imagery meets the minimum resolution thresholds required for effective discrimination.

The careful selection of UAV flight altitude is critical to ensuring both the quality of ecological data and the safety of operations. While lower altitudes can offer enhanced visual resolution conducive to detailed species identification, they simultaneously increase the risk of collision with aerial obstacles such as trees and tall riparian vegetation. One potential solution is to apply terrain-following flight paths based on pre-generated digital terrain models (DTMs), which can help maintain a constant relative altitude above ground while avoiding vegetation obstacles [45]. Equally important is the need to consider the potential disturbance caused by UAV-generated noise to waterbirds, as previous studies have shown that different species exhibit varying responses to UAV flight altitude [93]. For instance, surveys conducted at 15 m over gull colonies did not induce significant behavioral disturbances, suggesting that lower flight altitudes could be used for high-precision imaging in more disturbance-tolerant bird species [94]. Brisson-Curadeau et al. recommended using lightweight UAVs (≤2 kg) to survey nesting birds at distances of 20–25 m and encouraged baseline testing above 20 m to determine species-specific responses [95], which is consistent with the findings of this study. In general UAV-based ecological survey practices, flight altitude selection remains context-dependent and must carefully integrate sensor capabilities, GSD requirements, the size and distinctiveness of target features, operational efficiency considerations, and the need to minimize potential disturbance to wildlife [47,94,96].

In addition, species-specific behavioral and ecological traits considerably influence the detectability of waterbirds in UAV-based surveys. The three focal species in this study exhibit marked differences in habitat use and spatial behavior. Coots predominantly occupy open water zones and the margins of reed beds, feed primarily on aquatic vegetation, and exhibit strong territorial aggression, often excluding other waterbirds from preferred areas. Their nesting behavior typically involves constructing conspicuous floating nests in exposed locations. In contrast, moorhens, being omnivorous and more secretive, tend to forage and nest within dense riparian vegetation or grassland fringes, often relocating to peripheral microhabitats in response to interspecific competition. Ducks, differing from both, frequently nest at substantial distances from waterbodies, selecting dry, elevated zones with sufficient vegetative cover, such as riverbanks or semi-natural meadows. These behavioral divergences imply that if UAV imagery acquisition is limited to water surfaces or immediate shoreline zones, aerial surveys are likely to underestimate the presence and abundance of duck species due to their peripheral nesting preferences.

Habitat structural heterogeneity further complicates detection accuracy. In environments characterized by dense vegetation, tall herbaceous cover, or heterogeneous terrain, waterbird individuals may be obscured or visually blended into the background, reducing the performance of automated detection algorithms. Therefore, UAV-based survey protocols should be carefully adapted to the ecological context of the target species. Optimal strategies may include scheduling flights during periods of low solar glare and reduced avian activity to minimize shadow interference, expanding the survey footprint to encompass upland habitats such as floodplain meadows or embankments, and employing multi-angle imaging techniques to enhance object visibility. These adjustments are essential for improving detection rates across waterbird species with diverse nesting ecologies and concealment strategies.

4.3. Limitations and Future Research

Hardware limitations remain a critical challenge in UAV-based waterbird monitoring. This study relied solely on an RGB camera, which proved effective in open water areas but exhibited limitations in detecting birds in densely vegetated regions. Occlusion effects caused by tree branches, emergent vegetation, and shoreline structures hindered detectability, particularly for species with cryptic coloration or small body size. In addition, current methods rely heavily on manual image annotation, which is labor-intensive, time-consuming, and prone to observer bias. This study focused exclusively on the waterbird nesting season and lacks long-term monitoring data covering all seasonal variations.

Future research should consider using higher-resolution RGB cameras and combining them with other types of sensors. For instance, multispectral systems incorporating near-infrared (NIR) and thermal infrared cameras may assist in locating warm-bodied birds partially concealed by vegetation, although their relatively low spatial resolution limits their standalone effectiveness [97,98,99,100]. Additionally, AI-driven automation presents substantial potential for enhancing UAV-based waterbird monitoring [20,101]. Deep learning algorithms trained on large annotated datasets can automate species identification, significantly improving efficiency and consistency [101,102]. AI models can be integrated into UAV post-processing workflows to enable real-time object detection, reducing the time required for data analysis while improving detection accuracy. By integrating UAV remote sensing, AI-driven analysis, and ecological modeling, future waterbird monitoring efforts can achieve greater precision, efficiency, and scalability, ultimately contributing to more effective aquatic ecosystem conservation and management strategies [18,101,103].

5. Conclusions

This study demonstrates the effectiveness of UAV-derived high-resolution orthomosaics for quantifying waterbird abundance and elucidating habitat preferences in aquatic ecosystems. Results along the Lieve Canal revealed that herbaceous vegetation zones consistently supported the highest densities of waterbirds, underscoring the critical ecological role of low-disturbance, open vegetated habitats. In contrast, residential, industrial, and riparian tree zones showed reduced bird occupancy, highlighting the negative impact of human activities and unsuitable vegetation structures on waterbird distribution.

Validation against traditional ground surveys confirmed that UAV-based counts reliably captured overall waterbird abundance, though species detectability varied notably due to morphological traits and behavior. The Eurasian coot, with prominent visual markers, exhibited high detection accuracy, whereas the more cryptic moorhen presented significant identification challenges. The distinct terrestrial nesting habit of wild ducks necessitated survey area expansions beyond immediate water boundaries, emphasizing the need to consider species-specific ecological behaviors in survey design.

Spatial resolution emerged as a decisive factor influencing UAV detection performance. Finer ground sampling distances (GSD) markedly improved identification accuracy, particularly for smaller or cryptically colored species. Optimal operational practices include adjusting flight parameters based on species ecology, employing terrain-following flight paths, and carefully managing potential UAV disturbances to wildlife.

Current methodological limitations, such as sensor constraints, image occlusions, and labor-intensive manual annotation, highlight opportunities for methodological advancements. Future research integrating multispectral imaging and AI-driven analytics could substantially enhance detection precision, operational efficiency, and ecological understanding, ultimately supporting more effective conservation management in complex aquatic habitats.