Structural Health Monitoring of Anaerobic Lagoon Floating Covers Using UAV-Based LiDAR and Photogrammetry

Highlights

What are the main findings?

• Integrating UAV LiDAR with photogrammetry produced DEMs and orthomosaics approximately four times faster while maintaining centimetre-level accuracy.
• UAV-based monitoring of the anaerobic lagoon’s floating cover shows rapid uplift in the first two years, followed by lateral displacement and increased wrinkling associated with scum accumulation.

What is the implication of the main finding?

• This faster, accurate pipeline enables routine condition tracking and earlier interventions, improving asset reliability and safety.
• These diagnostics safeguard methane capture and odour control and provide a practical basis for digital twin and autonomous SHM at Melbourne Water’s Western Treatment Plant.

Abstract

There has been significant interest in deploying unmanned aerial vehicles (UAVs) for their ability to perform precise and rapid remote mapping and inspection of critical environmental assets for structural health monitoring. This case study investigates the use of UAV-based LiDAR and photogrammetry at Melbourne Water’s Western Treatment Plant (WTP) to routinely monitor high-density polyethylene floating covers on anaerobic lagoons. The proposed approach integrates LiDAR and photogrammetry data to enhance the accuracy and efficiency of generating digital elevation models (DEMs) and orthomosaics by leveraging the strengths of both methods. Specifically, the photogrammetric images were orthorectified onto LiDAR-derived DEMs as the projection plane to construct the corresponding orthomosaic. This method captures precise elevation points directly from LiDAR, forming a robust foundation dataset for DEM construction. This streamlines the workflow without compromising detail, as it eliminates the need for time-intensive photogrammetry processes, such as dense cloud and depth map generation. This integration accelerates dataset production by up to four times compared to photogrammetry alone, while achieving centimetre-level accuracy. The LiDAR-derived DEM achieved higher elevation accuracy with a root mean square error (RMSE) of 56.1 mm, while the photogrammetry-derived DEM achieved higher in-plane accuracy with an RMSE of up to 35.4 mm. An analysis of cover deformation revealed that the floating cover had elevated rapidly within the first two years post-installation before showing lateral displacement around the sixth year, which was also evident from a significant increase in wrinkling. This approach delivers valuable insights into cover condition that, in turn, clarifies scum accumulation and movement, thereby enhancing structural integrity management and supporting environmental sustainability at WTP by safeguarding methane-rich biogas for renewable-energy generation and controlling odours. The findings support the ongoing collaborative industry research between Monash University and Melbourne Water, aimed at achieving comprehensive structural and prognostic health assessments of these high-value assets.

1. Introduction

Over the past decade, UAV technology has advanced rapidly, with innovative development and applications across both commercial industries and research fields [1,2,3,4,5,6,7,8,9,10,11,12]. UAV-based remote sensing is particularly valued for its intrinsic operator safety, ability to access and scan otherwise inaccessible areas, and rapid, highly accurate approach to mapping and inspecting critical structures. Ongoing advances have also made UAVs more cost-effective and equipped them with a wide range of sophisticated sensing technologies, enhancing their versatility and effectiveness.

Consequently, there has been significant interest in deploying UAVs equipped with advanced sensing capabilities for structural health monitoring (SHM) of large and critical civil assets in infrastructure, mining and energy industries [8,12,13,14,15,16]. For instance, Zhao et al. [13] explored UAV-based photogrammetry as a cost-effective approach for emergency dam health monitoring, achieving centimetre-level accuracy with limited placement of ground control points (GCPs). Congress et al. [16] demonstrated that close-range UAV photogrammetry can provide rapid and accurate monitoring of sulphate heaving on road pavement infrastructure by analysing elevation and lateral displacements. Bolourian and Hammad [3] optimised flight path planning for bridge inspection by employing a LiDAR-equipped UAV to identify high-risk surface defect zones. Gaspari et al. [14] demonstrated a more efficient workflow for both data acquisition and processing by integrating UAV LiDAR and photogrammetry, achieving accuracies within 5 to 10 cm. Yan et al. [17] proposed an automated system for monitoring concrete cracks using UAVs, high-resolution images, and LiDAR data and employing a convolutional neural network to detect and quantify cracks at the pixel level with 85% accuracy. Collectively, these studies highlight that UAV technology is a crucial tool for advancing SHM and asset management of critical civil infrastructure.

Moreover, the adoption of UAV-based remote sensing technologies has grown rapidly across industries, particularly for inspecting and monitoring high-value assets in the advancing digital age. For instance, Chevron Corporation has deployed UAVs for methane emission detection and monitoring across all Gulf of Mexico platforms, marking a significant advancement from the prior reliance on handheld devices [9]. Another case study highlights the use of UAVs in refinery inspections to not only reduce maintenance-related risks but also enhance operational value by generating reliable records for future reference [10]. Since 2012, Shell has been actively employing and exploring the use of UAVs for routine inspections across multiple assets [12]. With the vast amount of data generated through these applications, Shell now aims to maximise the value of this data by transforming it into actionable insights that enhance informed decision-making via machine vision, advanced analytics, and digital twin applications. Furthermore, Oak Ridge National Laboratory’s Grid Communications and Security Group has been at the forefront of integrating artificial intelligence into the development and deployment of customised UAV swarms equipped with advanced sensing techniques, including visual and thermal imaging as well as LiDAR [18,19,20,21]. These UAVs have been deployed in various applications, such as the autonomous real-time inspection of damaged electrical infrastructure, i.e., powerlines and utility poles, providing real-time data to support operations and improve forest fire prevention strategies. Thus, UAV remote sensing is recognised as an indispensable tool for inspecting and assessing the integrity of critical assets across various industries, with its utilisation projected to expand further.

Melbourne Water’s Western Treatment Plant (WTP) in Werribee, Victoria, Australia, is a major wastewater treatment facility responsible for processing over half of Melbourne’s sewage, using low-cost, low-energy processes coupled with resource recovery [22]. Treatment comprises large anaerobic lagoon systems and activated sludge plants. These anaerobic lagoons are covered with high-density polyethylene (HDPE) membranes, as depicted in Figure 1. These membranes function as floating covers over the wastewater, maintaining an anaerobic environment crucial for bacterial digestion and capturing methane-rich biogas, which is harvested for energy generation that surpasses the plant’s energy needs, as well as suppressing obnoxious odours and greenhouse emissions [22,23]. Because influent sewage to the anaerobic section is unscreened and untreated, solidified scum matter can accumulate, creating large scumbergs that impose buoyant and contact loads on the cover, causing vertical uplift of approximately one metre and lateral displacement. Scum accumulation can also disrupt biogas collection channels integrated within the cover, thereby diminishing renewable energy efficiency. Effective asset management of this critical asset is therefore both an operational and environmental imperative, requiring timely monitoring to ensure safe and efficient operation.

The WTP is a world leader in environmentally sustainable, cost-effective, and energy-efficient sewage treatment, and it supplies recycled water for non-potable use. Melbourne Water is proactively pursuing leading-edge projects aligned with the United Nations Sustainable Development Goals and developing innovative approaches to monitor the structural integrity and performance of these critical assets. Resource recovery at WTP includes recycled water, which is used to irrigate on-site farmland and wetlands, and the nearby Werribee Irrigation District, and biosolids, positioning WTP as a key asset in Melbourne’s circular economy with a stated pathway to net-zero emissions by 2030 [22]. The site is also an internationally recognised Ramsar wetland and major bird habitat, underscoring the need to minimise odours and fugitive emissions. Within this framework, maintaining cover integrity is both an operational and environmental necessity.

Since 2016, collaborative industry research projects between Monash University and Melbourne Water have made innovative use of UAV-based remote sensing to monitor the floating covers at WTP. The overall goal is to achieve comprehensive structural and prognostic health assessments aligned with digital twin concepts [11]. Initially, photogrammetry-based UAV surveillance monitoring was implemented at the WTP lagoons, providing a low-cost, rapid assessment that offered a safer and less labour-intensive alternative to on-site walk inspections. A significant and immediate benefit was the high-resolution imagery, which provided engineers with a convenient tool for detecting surface deformations like the formation of creases (folds and wrinkles) over time, allowing for monitoring of the condition and movement of the cover [23,24,25]. Research efforts have focused on optimising UAV flight parameters for accurate 3D models and time-lapse monitoring of elevation changes, scum accumulation beneath the cover, and movement of the floating cover [11,26,27]. Furthermore, on-site investigations conducted between 2019 and 2020 [26] identified a strong correlation between scum depth and above-water cover elevation derived from both 3D modelling and laser surveying, providing a basis for further analysis. In addition to ensuring structural performance, effective monitoring of these anaerobic covers plays a vital role in environmental protection by preventing uncontrolled methane release, maintaining odour control, and safeguarding nearby water bodies and habitats, supporting environmental and biodiversity outcomes at WTP.

Furthermore, the research has empowered the industry by improving maintenance scheduling, facilitating more informed decision-making, and ensuring the prolonged and safe operation of the covers [23,25,28]. For example, 3D models of the 55E lagoon’s cover supported further verification of the restraining cables’ lengths. These models also supported the identification of excessive scum and sludge accumulations, which had the potential to damage the anaerobic covers and led to the partial cover replacement and desludging project in 2020. Additionally, 3D models have supported the analysis of scum growth and the performance of the 25 W anaerobic lagoon, helping to mitigate further decline and reduce the risk of process failure [28].

Although UAV-based photogrammetry offers significant benefits, generating 3D models through this method can be very time-consuming and often requires further efforts to render them usable [15,29,30,31]. Prior studies have incorporated additional inspection data, such as LiDAR, to further enhance their geospatial datasets [32]. These enhancements include establishing control datums and georeferencing methods via LiDAR data while leveraging RGB imagery photogrammetry method [32,33,34] and using LiDAR data for image orthorectification [35,36,37]. Liu et al. [37] demonstrated that orthorectified images obtained using LiDAR-derived data yield significantly less error than those derived from photogrammetry alone. Furthermore, the fusion of multiple data sources has been utilised to enhance analysis and modelling tasks such as classification [38,39,40,41]. For instance, Zhang et al. [39] fused LiDAR and photogrammetry point clouds to obtain colour information for visualisation and performed 3D building extraction using U-Net model segmentation. Their approach achieved higher performance metrics than methods using individual data sources. Similarly, Sankey et al. [41] classified soil nutrient and plant species changes with an overall accuracy of 87% using UAV hyperspectral and LiDAR fusion, compared to an accuracy of 71% when using only UAV hyperspectral data.

Recently, LiDAR technology has been incorporated into UAV surveillance operations to enhance asset management at WTP. This case study investigates the use of LiDAR and photogrammetry data for monitoring the movement of floating covers on the 25W anaerobic lagoon. The proposed approach combines the data and strengths of both technologies, enabling more efficient 3D modelling and analysis for monitoring cover integrity over time. This paper first details the integration of LiDAR-derived point clouds and photogrammetric images to produce geospatial data, specifically digital elevation models (DEMs) and orthomosaics. The development and accuracy of these geospatial data are compared against those produced using the previous photogrammetry method. The second part of the investigation focuses on analysing the displacement of the floating cover over time, with a discussion of potential implications for the 25W anaerobic lagoon at the Melbourne Water WTP. The outcomes of this study streamline the development of readily usable geospatial data for immediate analysis, revealing characteristics of scum accumulation, while supporting the overarching goal of achieving real-time, autonomous monitoring of asset structural integrity in the digital era.

2. Methods

2.1. Overview of Workflow

The overall workflow is schematised in Figure 2, which highlights stages from asset preparation and UAV configuration through to processing of UAV-based geospatial data for SHM. In brief, the approach utilised both LiDAR and photogrammetry, hereafter referred to as the integrated LiDAR–photogrammetry (LP) approach, by orthorectifying photogrammetric images onto DEMs derived from LiDAR point clouds. The geospatial data were processed following the steps below.

An overview of the algorithmic processing workflow is shown in Figure 3. The workflow begins with UAV-derived DEM and orthomosaic acquisition, followed by preprocessing and coordinate alignment. To assess spatial accuracy, UAV-derived outputs were compared against ground-survey laser checkpoints (CPs) using a set of statistical metrics. DEM quality was examined through residual analysis and filtering methods. Natural features and artificial markers were subsequently extracted and matched using normalised cross-correlation, with manual inspection and parameter adjustment applied to weak matches. Displacement analysis was then carried out through baseline and current-state subtraction, and results are visualised as interpolated displacement vector fields and elevation maps. The following subsections provide detailed descriptions of each stage of the workflow.

The methodological innovation lies in the anchoring of RGB-derived in-plane texture to LiDAR-based elevation data through orthorectification, thereby allowing coherent analysis in both planimetric and altimetric domains. By leveraging correlation-based feature tracking with correlation-quality gating, the method enables deterministic and reproducible mapping of surface displacements. Furthermore, residual masks derived from DEM differencing enable the spatial quantification of surface wrinkling, offering a novel metric for assessing scum accumulation and mechanical behaviour of floating covers in WTP.

2.2. UAV Data Acquisition and Flight Setup

Recently, an affordable UAV-equipped LiDAR system, DJI Zenmuse L1, became available, offering enhanced capabilities for high-resolution 3D mapping and data acquisition. In 2022, the DJI Matrice M300, equipped with the Zenmuse L1 LiDAR camera, was deployed to scan the anaerobic lagoons at WTP, replacing the previous bespoke UAV systems documented in prior works [27,42]. The Zenmuse L1 is mounted on a stabilised 3-axis gimbal and includes an integrated RTK GNSS and IMU system for precise georeferencing of point clouds. Its RGB mapping camera, the EP800 optical component, features a 20 MP 1-inch CMOS sensor with a global shutter and an 8.8 mm focal length. According to the manufacturer’s specifications, the LiDAR module system achieves a horizontal accuracy of 10 cm and a vertical accuracy of 5 cm when flown at an altitude of 50 m and a speed of 10 m/s. Previous studies [14,43] conducted comprehensive tests of the DJI Zenmuse L1 sensor, demonstrating that, after systematic transformation corrections, the positioning accuracy of all directions exceeds the manufacturer’s specifications.

For the present investigation, the flight operation at the 25W anaerobic lagoon was conducted at an altitude of 70 m using a DJI D-RTK 2 High Precision GNSS Mobile Station to correct the geotagged locations of images and LiDAR to the UAV in real-time. DJI Pilot was used to manage the UAV flight plan and configurations. As shown in

Table 1. DJI Pilot flight and LiDAR parameter settings.

Flight Parameters		LiDAR Specific Settings
Altitude	70 metres	Return Mode	Dual
GSD	1.91 cm/pixel	Sampling Rate	240 kHz
Point cloud density	748 points/m2	Scanning Mode	Repetitive
Side Overlap (LiDAR)	70%	RGB Colouring	On
Side Overlap (Visible)	76%
Forward Overlap (Visible)	70%
Flight Speed	8.3 m/s

, the flights maintained a ground sampling distance (GSD) of 1.91 cm/pixel, resulting in a point cloud density of 748 points/m², with LiDAR and image overlaps of 70%. A total of 469 images, with a pixel resolution of 5472 by 3648 (pixel size of 2.41 µm), were captured, and the data collection duration for the flight was approximately 30 min. Furthermore, a total of six GCPs, denoted as GCP1 to GCP6, pre-surveyed in the concrete area around the 25W anaerobic lagoon, were used for calibration purposes. In this investigation, both the LiDAR-derived DEM with an RTK station and one GCP (labelled GCP6) and the photogrammetry-derived DEM using all six GCPs were examined (refer to Figure 4), using the same DJI Matrice M300 UAV equipped with the DJI Zenmuse L1 for consistency.

2.3. UAV LiDAR and Photogrammetry Approach

2.3.1. Flight Surveys and Dataset

Seven geospatial datasets acquired from December 2018 to November 2023 were processed using Agisoft Metashape Professional 1.7.2 and MATLAB R2023b using built-in functions from the Image Processing and Computer Vision toolboxes. The geospatial data collected before 2022 were obtained via UAV-photogrammetry [27,42], while those collected afterwards employed the integrated UAV-LP method. The geospatial data were transformed into a local Cartesian coordinate system and, for consistency, cropped to a size of 9251 by 24,001 pixels, with a resampled resolution of 2 cm/pixel using bicubic interpolation. It is worth noting that the initial geospatial data of the 25 W anaerobic lagoon, captured in December 2018, were the very first of their kind and had not been optimised in terms of image quality and flight settings, resulting in the lowest resolution of approximately 5 cm/pixel. Furthermore, the floating cover for the 25 W anaerobic lagoon, including biogas extraction, was installed and commissioned in April 2017. At that time, it rested directly on the water surface (cover at zero above-water elevation), as no scum was present. Across periods, feature matching is performed on the orthorectified RGB orthomosaic, elevations are sampled from LiDAR-derived DEMs where available and from photogrammetry-derived DEMs otherwise. All UAV geospatial data are transformed into a cover-referenced local frame, zeroed to water level, and resampled to a common resolution to ensure comparability across periods.

2.3.2. Photogrammetry Method

Image alignment, including all associated metadata (e.g., GPS location and camera settings), and the generation of photogrammetry-derived DEMs were performed using Agisoft. For the comparison study, all settings, including image alignment (with a key point limit of 100,000 and a tie point limit of 40,000) and dense cloud configuration (using Mild Depth Filtering), were set to the highest quality (“Ultra high”) to achieve a resolution of less than 3 cm/pixel for both the DEMs and the orthomosaics. Additionally, the second-highest quality (“High”) settings were also used for comparison. This resolution also provides sufficient quality for cover movement analysis in the second part of the study. As a standard practice, repeated validation and calibration steps were employed to mitigate systematic distortion errors, ensuring accuracy, consistency and reliability in the results. The processing was conducted on a dedicated workstation equipped with an Intel i7-9700 CPU (3.00 GHz), 128 GB of RAM, and an NVIDIA Quadro P4000 GPU with 8 GB of VRAM.

2.3.3. Integrated LiDAR and Photogrammetry Method

An integrated LP approach was pursued, as follows:

The acquired LiDAR point cloud was initially processed using DJI Terra Software Version 3.6, with the WGS 84/UTM Zone 55S coordinate system (EPSG:32755). The processed LiDAR point cloud data were then imported into Agisoft, where LiDAR-derived DEMs were generated using the same highest-quality settings described for the photogrammetry approach. Subsequently, the images were aligned and orthorectified using the LiDAR-derived DEMs as the projection plane to construct the corresponding orthomosaic, as indicated in Figure 2. This approach enhances visual quality and interpretability by providing high-resolution imagery that accurately captures details and colours while maintaining precise terrain models. Conceptually, LP constrains elevation with LiDAR and preserves in-plane texture with RGB, yielding a geometry-anchored orthomosaic suited to unified in-plane and out-of-plane analysis. Furthermore, performance timings and comparative outputs of each method are reported.

2.4. Geospatial Data Accuracy Assessment

To quantify positional accuracy and alignment quality of the geospatial data, the root mean square error (RMSE), refer to Equation (1), was used as the primary error metric throughout this study. It was applied in two contexts (i) to evaluate the accuracy of geospatial data against on-site laser survey CP, and (ii) to assess the local coordinate transformation by verifying the alignment of predefined static features (Section 2.5.1). The RMSE is defined as:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{‖{\mathit{p}}_{\mathit{i}}-{\mathit{p}}_{\mathit{i}}^{\prime }‖}^2}$$

(1)

where ${\mathit{p}}_{\mathit{i}}$ is the reference coordinate (e.g., laser survey measurement or predefined feature), ${\mathit{p}}_{\mathit{i}}^{\prime }$ is the corresponding derived coordinate and $n$ is the number of sample.

2.4.1. Porthole Checkpoints

The laser survey was conducted independently by the Melbourne Water Geospatial Surveyor team using a Leica TM60, with an on-site survey control accuracy within 5 mm, resulting in an overall precision of ±15 mm at a 95% confidence level. Each CP was marked with a self-adhesive RS90R Datum Target Level Marker, and during the survey, a target height was applied to measure the surface of the porthole. Coordinate datum measurements were conducted using the GDA 94/MGA Zone 55 and AHD systems. The RMSE was then calculated to determine the difference between the coordinates of 10 CPs located on the portholes, denoted as 25W-01 to 25W-10, in the geospatial data and the on-site laser survey measurement results (see Figure 5). To assess accuracy stability, repeated k-fold cross-validation (CV) was performed on the 10 CPs (k = 3 and 50 repeats), which yields 3–4 CPs per test fold and, through repetition, smooths fold-to-fold variability. In each partition, one fold was held out, and RMSE was computed on the held-out CPs. The 150 held-out RMSEs (3 folds of 50 repeats) were then aggregated to obtain the CV mean RMSE and its standard deviation (σ) across folds for Easting, Northing, and elevation.

2.4.2. DEM Noise Quality

DEMs are known to be susceptible to various types of noise and artifacts, including random pixel-level distortions such as salt-and-pepper noise, speckle patterns inherent in the data, background fluctuations caused by environmental, sensor, or equipment-related factors and systematic errors introduced during data processing [15,44,45]. In this study, the quality of the DEM derived from both the LiDAR point cloud and photogrammetry was evaluated for an area of interest at CP 25W-07, as shown in Figure 4. This CP was selected because it contains distinct features, such as wrinkles and man-made structures, that are representative of the cover surface and allow for a meaningful visual and statistical assessment of DEM noise characteristics.

Residual analysis was performed to assess the noise within this area. However, as achieving a perfectly clean DEM is unattainable, a 2D median filter was employed to suppress isolated pixel noise:

$${\mathrm{D}\mathrm{E}\mathrm{M}}_{\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{d}}\left(i,j\right)=\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\left\{\mathrm{D}\mathrm{E}\mathrm{M}\left(i+m,j+n\right)| m,n \in \left[-k, k\right]\right\}$$

(2)

where $\mathrm{D}\mathrm{E}\mathrm{M}\left(i,j\right)$ denotes the elevation value at pixel $\left(i,j\right)$, and the median is computed over an $\left(2k+1\right)$ by $\left(2k+1\right)$ pixel neighbourhood centred at $\left(i,j\right)$. A high-pass DEM was then obtained by subtracting the filtered DEM from the original:

$${\mathrm{D}\mathrm{E}\mathrm{M}}_{high-pass}\left(i,j\right)={\mathrm{D}\mathrm{E}\mathrm{M}}_{\mathrm{o}\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}\left(i,j\right)-{\mathrm{D}\mathrm{E}\mathrm{M}}_{\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{d}}\left(i,j\right)$$

(3)

For noise analysis, Equation (2) was applied with $k=5$ (11-by-11-pixel window). The resulting high-pass DEM in this case is referred to as the residual DEM.

2.4.3. Visualisation Enhancement of Wrinkle Profile

Wrinkles and folds on the floating covers are recognised by Melbourne Water engineers and technicians as vital indicators of high-stress concentrations, which may compromise the structural integrity of the cover and reduce biogas capture efficiency or premature material degradation [23,24,25]. Traditionally, these features have been monitored through on-site visual inspection and photography. While effective at a local scale, such approaches are labour-intensive and lack consistency over time. In contrast, geospatial data enables a proactive, full-field characterisation of wrinkles and their evolution, providing enhanced insight for decision-making and long-term asset management.

For wrinkle visualisation, the same median filtering and high-pass operations defined in Equations (2) and (3), respectively, were applied with a larger kernel size of $k=50$, corresponding to 100-by-100 pixels. This configuration suppresses long-wavelength undulations while isolating wrinkle-scaled ridges. The resulting high-frequency DEM was then binarised using a threshold ($\tau$):

$$B\left(i,j\right)=\left\{\begin{array}{c}1, DE{M}_{high-pass}\left(i,j\right)\ge \tau ,\hfill \\ 0, otherwise.\hfill \end{array}\right.$$

(4)

Morphological area opening was applied to remove all connected components with fewer than $\alpha$ pixels. Let $\mathbb{C}=\left\{ C_1, C_2, \dots , C_k \right\}$ be the set of 8-connected components in $B$, with $\left|C_k\right|$ denoting the pixel area of components $C_k$. The clean wrinkle mask was then defined as:

$$B_{clean}\left(i,j\right)=\left\{\begin{array}{c}1, \left(i,j\right) \in C_k for some C_k\in C with \left|C_k\right|\ge \alpha ,\hfill \\ 0, otherwise.\hfill \end{array}\right.$$

(5)

Additionally, features such as trapped rainwater were identified using a bespoke image segmentation approach, which applied unsupervised machine learning to filter characteristics of the anaerobic lagoon’s floating cover [46]. In summary, this process involved employing an unsupervised k-means clustering algorithm on a stacked 4D array composed of the DEM elevation and the RGB channels of the associated orthomosaic. The optimal number of clusters was determined based on a within- and between-cluster variance criterion to ensure the resulting groups were interpretable and meaningful. Through inspection, the cluster associated with the trapped rainwater was selected and the remaining clusters were removed. Further details of this procedure can be found in the reference [46].

2.5. Relative In-Plane and Out-of-Plane Displacements Estimation

An image-matching strategy based on normalised cross-correlation (NCC) with spatial constraints was employed for two purposes: (i) to perform local coordinate system transformation by aligning static features across datasets, and (ii) to identify and track natural and artificial square markers for displacement analysis across seven sets of geospatial data. This strategy was selected for its practical suitability in the WTP environment with minimal tuning, deterministic and fast execution on field hardware, and a correlation score that supports straightforward quality control, making it well suited to routine industrial surveys.

The NCC between a template image $T\left(i,j\right)$ and a search region $I\left(i,j\right)$ was defined as:

$$NCC\left(u,v\right)={\displaystyle \frac{{\sum }_{i,j}\left(T\left(i,j\right)-\overline{T}\right)\left(I\left(i+u,j+v\right)-{\overline{I}}_{u,v}\right)}{\sqrt{{\sum }_{i,j}{\left(T\left(i,j\right)-\overline{T}\right)}^2·{\sum }_{i,j}{\left(I\left(i+u,j+v\right)-{\overline{I}}_{u,v}\right)}^2·}}}$$

(6)

where $\left(u,v\right)$ denotes the displacement offset, $T\left(i,j\right)$ is the template window, $I\left(i,j\right)$ is the search window, and $\overline{T}$ and ${\overline{I}}_{u,v}$ are the mean intensity values of the respective regions. The maximum correlation strength indicates the best match location. The method involves searching for a specified sub-image within a limited search region, which improves the speed and accuracy of the matching process while remaining robust to variations in illumination and contrast.

2.5.1. Local Coordinate System Transformation

A local coordinate system clarifies the visualisation of cover movement by referencing results to the cover itself rather than to absolute survey coordinates. Expressing displacements in this intuitive local frame makes the analysis more interpretable for engineers and technicians and directly supports asset-management decisions. The local frame is established by using static cover features as fixed anchors for image-based matching.

The procedure for local coordinate system transformation is as follows: the orthomosaic was first converted to grayscale. Seventeen static features were predefined based on their locations as of May 2023. Each feature, centred within a 100-by-100 pixel (2 m-by-2 m) sub-image (see Figure 6 and refer to

Table 2. Template and search window sizes for feature matching.

Feature Type	Template Window Size (Pixels)	Search Window Size (Pixels)
Static features	100 × 100	160 × 160
Natural features	200 × 200	100 × 100
Artificial square marker features	10 × 10	100 × 100

), was located by identifying the maximum NCC strength within a localised search region of 3.2 m-by-3.2 m, centred on the predefined coordinates. Using these matched feature coordinates, a 2D similarity transformation was estimated to align the datasets into a common local coordinate system, using MATLAB’s built-in function fitgeotform2d.

The similarity transformation maps each moving feature point $\left(x_k,y_k\right)$ to the predicted fixed point $\left({\widehat{x}}_k^{\prime },{\widehat{y}}_k^{\prime }\right)$ as:

$$\left[\begin{array}{c}{\widehat{x}}_k^{\prime }\\ {\widehat{y}}_k^{\prime }\end{array}\right]=s \left[\begin{array}{cc}\cos \theta & -\sin \theta \\ \sin \theta & \cos \theta \end{array}\right]\left[\begin{array}{c}{x}_k\\ y_k\end{array}\right]+\left[\begin{array}{c}{t}_x\\ t_y\end{array}\right],$$

(7)

where $s$ is a uniform scale factor, $\theta$ is the rotation angle, and $\left(t_x,t_y\right)$ are translations.

The translation parameters $\left(s,\theta , t_x,t_y\right)$ were obtained by minimising the sum of squared residuals across all $N=17$ static feature pairs:

$$\underset{s,\theta , t_x,t_y}{\min }\sum _{k=1}^N{‖\left[\begin{array}{c}{x}_k^{\prime }\\ y_k^{\prime }\end{array}\right]-\left[\begin{array}{c}{\widehat{x}}_k^{\prime }\\ {\widehat{y}}_k^{\prime }\end{array}\right]‖}^2$$

The alignment error was quantified using the RMSE, where in Equation (1), ${\mathit{p}}_{\mathit{i}}={\left(x_k^{\prime },y_k^{\prime }\right)}^{\top }$ is the $k$-th observed fixed (static feature) coordinate and ${\mathit{p}}_{\mathit{i}}^{\mathit{’}}={\left({\widehat{x}}_k^{\prime },{\widehat{y}}_k^{\prime }\right)}^{\top }$ is the $k$-th predicted transformed coordinate.

2.5.2. Tracking of Natural and Artificial Features

In the area of interest, the Eastern Section was divided into four segments, denoted ES1 to ES4, and the Eastern Segments and the three Eastern Flotation levels (top, centre, and bottom), as well as the two Bowtie Flotations (East and West), were inspected for displacement analysis, see Figure 6. A total of 385 features on the floating cover were analysed, comprising 240 natural features (such as portholes, welding lines, and flotation devices) and 145 10 cm square markers. These square markers were placed in the Eastern Section starting in 2021 and were fully in place by May 2023. The area of interest encompasses 325 features with nearest neighbour Euclidean distances ranging from 36.1 pixels (0.72 m) to 1432.4 pixels (28.6 m), with a mean distance of 328.1 pixels (6.6 m). For geospatial data collected before May 2023, only 240 features were considered for analysis, with 180 of those features located in the area of interest. The nearest neighbour Euclidean distances for these features ranged from 49.0 pixels (1.0 m) to 1815.5 pixels (36.3 m), with a mean distance of 435.2 pixels (8.7 m).

After coordinate alignment, relative displacements of the floating cover were quantified using both natural features and artificial 10 cm square markers placed in the Eastern Section. Similarly, the image matching using NCC for displacement analysis is as follows: The orthomosaic was converted to grayscale. Refer to Table 2, natural features were centred in 200-by-200 pixel (4 m-by-4 m) sub-images and square markers were centred in 10-by-10 pixel (20 cm-by-20 cm) sub-images, with their reference coordinates predefined from the May 2023 geospatial dataset. Each feature was then located by identifying based on their maximum NCC strength within a 2D search region extending ±50 pixels (1 m) from the centre of its predefined coordinates. Detected features with a NCC strength below 0.6 were flagged as potential errors, prompting manual user inspection. The procedure was then modified by adjusting the search window, sub-image size, or using a sub-image from a different time instance to improve the accuracy of feature localisation. Features that could not be located were either manually located or excluded from the analysis, such as those submerged in trapped rainwater or obscured by excessive debris.

The identified features at each of the seven geospatial datasets were then used to determine the relative in-plane and out-of-plane displacements based on the earliest and latest geospatial data. This was done by baseline and current-state subtraction approaches to explore the cumulative and global changes since the initial state (December 2018) and to explore the most recent changes in the floating covers, providing more informative insights into the effects of biogas pockets and trapped rainwater leading to the current state (November 2023), respectively. Baseline subtraction used only the 145 natural markers present in December 2018 across all cases to ensure consistency. For the current-state subtraction, the analysis uses all available markers at each time instance, matching the number of markers to the data available at that specific time (e.g., all 385 markers from May 2023 to November 2023, 276 markers in June 2022, and 240 markers from December 2018 to November 2020).

For visualisation purposes, the relative magnitude in-plane and out-of-plane displacement data were linearly interpolated onto a uniform grid with a pixel size of 50. Additionally, the relative in-plane displacements were interpolated onto a uniform grid with a pixel size of 500, resulting in an in-plane vector map indicating the directionality. Additionally, to facilitate the explanation of the movement of floating covers, the cardinal directions—North, South, East, and West—are designated as N, S, E, and W, respectively, relative to the local coordinate system (see Figure 6). The results and discussion focus primarily on the maximum above-water elevation and relative displacements of the floating cover, particularly in the Eastern Section (i.e., segments and flotations) and the Bowtie Flotations.

2.6. Theoretical and Technological Innovations

This work introduces a UAV-based geospatial workflow that integrates LiDAR and photogrammetry to enable coherent, full-field displacement analysis and tracking at the floating-cover asset scale. The key advances are:

• Geometry-anchored fusion. High-resolution RGB imagery is orthorectified onto LiDAR-derived DEMs, producing a geometry-anchored orthomosaic in which LiDAR constrains out-of-plane elevation while image alignment preserves in-plane texture. This bypasses photogrammetric dense-depth reconstruction and enables unified analysis in all directions.
• Wrinkle mapping from high-pass DEM. A median-filter residual (high-pass DEM) isolates wrinkle features. Thresholding and morphological area opening yield a reproducible wrinkle mask that serves as a condition indicator related to local stress concentration and cover-scum behaviour.
• Cover-referenced local frame. Displacements are expressed in a local coordinate system tied to the cover, constructed by NCC matching of static features and estimation of a 2-D similarity transformation. This renders results invariant to global translation/rotation/scale across surveys and improves operational interpretability.
• Deterministic feature tracking with quality control. NCC within bounded search windows is used to track natural features and artificial markers. Ambiguous matches are flagged for manual inspection under a predefined protocol.

Collectively, these elements provide a technology-integrated method that anchors geometry to LiDAR, preserves texture for displacement analyses, and yields reproducible displacement and condition indicators suitable for routine industrial surveys.

3. Results

3.1. Generation of Geospatial Data

The LiDAR and photogrammetry processing duration for each stage was recorded, and the resulting geospatial data from each technique were analysed and compared. The orthomosaics were also visually inspected for inaccuracies in the orthorectification. Orthomosaic seamline editing was conducted to ensure a more accurate representation of the geospatial image and eliminate incorrect image orthorectification. This was achieved by assigning multiple suitable images, excluding inappropriate ones in a selected region, and then updating the orthomosaic accordingly.

As shown in

Table 3. Information on geospatial data generation using photogrammetry and LP methods.

Photogrammetry Method					Integrated LiDAR-Photogrammetry Method
Processing Stage	Duration	Resolution	Duration	Resolution	Processing Stage	Duration	Resolution
Quality Setting	Highest		High		Quality Setting	Highest
Align Photos	1 h	-	1 h	-	Align Photos	1 h	-
					Point Cloud Processing (via DJI Terra)	<30 min	-
Dense Clouds (Depth Maps) (Dense Point Cloud)	~10 h (3 h) (7 h)	506,531,251 pts	~4 h (1 h) (3 h)	129,383,883 pts	Dense Clouds	~1 h	207,484,387 pts
DEM	<5 min	1.93cm/pixel	<5 min	3.85 cm/pixel	DEM	<5 min	2.90 cm/pixel
Orthomosaic	<30 min	1.93cm/pixel	<30 min	1.93 cm/pixel	Orthomosaic	<10 min	2.13 cm/pixel
Total Duration	~12 h		~5 h		Total Duration	~3 h

, the integration of LiDAR with photogrammetry in DEM generation accelerates the process, making it approximately four times faster than using photogrammetry alone at the highest settings and nearly twice as fast as photogrammetry at the high-quality setting. This speed advantage arises primarily from LiDAR’s direct elevation capture, which eliminates the need for dense cloud and depth map generation. These computationally intensive steps represent a major time demand, accounting for approximately 80% of the photogrammetry workflow. By leveraging both methods, the process benefits from photogrammetry’s visual detail and LiDAR’s efficient, direct elevation data, thereby optimising the DEM and orthomosaic creation pipeline. Specifically, the LiDAR point cloud serves as the source data for constructing the DEM due to its speed and efficiency. This DEM is then used in conjunction with aligned and orthorectified photogrammetric images to produce the associated orthomosaic, delivering a detailed and accurate representation of the surface. Hence, this approach bypasses the computationally complex procedures required for DEM generation using photogrammetry, thereby reducing the overall processing time.

With the highest quality setting, photogrammetry-derived DEMs yield a superior resolution of 1.93 cm/pixel compared to LiDAR-derived DEMs, which achieve 2.90 cm/pixel, due to their higher point density. At the high-quality setting, the photogrammetry-derived DEM had the lowest resolution of 3.85 cm/pixel. The quality of the orthomosaics remained at 1.93 cm/pixel, with a relatively reduced resolution of 2.13 cm/pixel for the LP method.

3.2. Accuracy of LiDAR and Photogrammetry Geospatial Data

Overall, both LP and photogrammetry-derived geospatial data achieve similar centimetre-level accuracy in all directions when compared to on-site laser survey measurements. Specifically, LP produced a more accurate elevation value, with an RMSE of 56.1 mm compared to the photogrammetry-derived RMSE of 69.5 mm. However, photogrammetry-derived geospatial data exhibited more accurate in-plane components, resulting in RMSE values for Easting and Northing of 35.4 mm and 24.0 mm, respectively, compared to the LP-derived data, which had RMSE values of 64.4 mm and 62.1 mm, respectively. In terms of accuracy stability, CV on the LP-derived CP errors yielded mean held-out RMSEs of 61.6 mm (Easting), 59.6 mm (Northing), and 45.4 mm (Elevation), with fold-to-fold σ of 18.9, 17.4, and 32.1 mm, respectively. These are consistent with the full-sample LP RMSEs, indicating that no single CP drives the accuracy result. Performing the same analysis on the photogrammetry-derived errors gave mean held-out RMSEs of 34.7 mm (Easting), 23.2 mm (Northing), and 67.0 mm (Elevation), with σ of 7.4, 6.0, and 19.4 mm, closely matching the full-sample photogrammetry RMSEs (35.4, 24.0, 69.5 mm). The larger elevation σ (i.e., LP-derived geospatial data at 25W07), suggests greater local variability in elevation (discussed in the next section), while the overall pattern remains.

3.3. LiDAR and Photogrammetry-Derived DEM Noise Quality

As shown in Figure 7, LiDAR-derived and photogrammetry-derived DEMs successfully capture large features, such as portholes and flotations, as well as wrinkling. The relatively small kernel size was empirically chosen because it effectively suppresses isolated pixel-level noise, such as salt-and-pepper noise, while preserving local surface variations that are relevant to SHM. Notably, photogrammetry-derived DEM, in particular, resolved finer textural details, including cables and ballast. By contrast, LiDAR DEMs have greater noise, characterised by impulse-like artifacts, which is primarily due to the inherent limitations in laser-based data acquisition methods [44], such as the scattering of lasers due to environmental factors. At CP 25W-07, for instance, dirt and surface water were observed, which likely contributed to the higher elevation error reported in

Table 4. Errors of integrated LP-derived and photogrammetry-derived geospatial data relative to on-site laser survey checkpoints.

	LP-Derived Geospatial Data Error (mm)			Photogrammetry-Derived Geospatial Data Error (mm)
Porthole	Easting	Northing	Elevation	Easting	Northing	Elevation
25W-01	−39.2	110.9	−17.1	−25.7	25.0	33.7
25W-02	80.2	45.2	−18.1	39.9	−22.7	109.9
25W-03	48.0	−20.7	20.4	49.6	−20.0	39.9
25W-04	5.1	−60.8	19.2	31.0	23.6	46.6
25W-05	−67.3	−61.3	3.1	24.1	46.4	−29.4
25W-06	−11.6	−10.4	1.0	58.2	−5.2	−40.8
25W-07	11.9	−7.4	−158.5	35.6	−18.8	−64.9
25W-08	−69.9	80.0	14.1	−39.5	31.8	15.3
25W-09	−131.1	−97.7	68.5	4.2	0.9	−87.3
25W-10	−65.2	−16.2	−6.8	7.4	10.8	−129.3
RMSE	64.4	62.1	56.1	35.4	24.0	69.5
CV mean RMSE	61.63	59.56	45.42	34.68	23.22	67.03
CV σ RMSE	18.93	17.44	32.15	7.39	5.98	19.42

.

In

Table 5. Statistic on the noise from UAV-LiDAR and photogrammetry-derived DEM.

	Median	Q1	Q3
LiDAR-derived DEM	7.3 mm	2.3 mm	14.4 mm
Photogrammetry-derived DEM	2.0 mm	0.8 mm	3.7 mm

, statistical distribution, including quartiles, was analysed to evaluate the spatial variability of noise within the 25W-07 region. The results indicate that the LiDAR-derived DEM errors show deviations ranging from 2.3 mm to 14.4 mm, whereas the photogrammetry-derived DEM errors show deviations ranging from 0.8 mm to 3.7 mm. Residuals exceeding 15 mm constitute a mean proportion of 6.4% in the LiDAR-derived DEM, compared to only 0.8% in the photogrammetry-derived DEM. Additionally, the relative mean proportion of error within the region is 29.5% for the LiDAR-derived DEM, indicating a moderate dispersion of errors and reflecting variability in their occurrence. In contrast, the photogrammetry-derived DEM shows a lower mean proportion of error at 20.8%.

In both cases, DEM noise was strongly associated with surface features on the membrane, particularly in areas with pronounced elevation change.

3.4. Wrinkling and Ridges on the Floating Cover

A comparison of wrinkles and small spatial features between the December 2018 and November 2023 DEMs is presented in Figure 8. The thresholds $\tau = 15 \mathrm{c}\mathrm{m}$ and $\alpha =10$ were empirically selected to highlight significant wrinkles and ridges while minimising background noise. The appearance of wrinkles in December 2018 suggests the floating cover has already been displaced in the in-plane direction perpendicular to the ridges in wrinkles, specifically toward the NW direction in ES1 and toward the SW direction in ES3.

From December 2018 to November 2023, the number of wrinkles has significantly increased across all sections, particularly in ES2 and ES3 near the bowtie section, at the top of the bowtie section, and in the Western Section. Figure 8b illustrates that the floating cover has been further displaced, with the previous wrinkles shifting NW, as indicated by the displacement of the porthole, ballast, and flotation devices in ES1. By November 2023, the emergence of new vertical wrinkles in ES2 and ES3 and in the Western Section indicates a westerly displacement of the cover.

Moreover, the appearance of ridges has also increased substantially, particularly in areas of high elevation on the east side of the Eastern Section. These changes can be attributed to scum accumulation leading to contact with the floating cover, causing the cover to conform to the surface of the scum. Additionally, trapped rainwater has formed localised sink areas, leading to the formation of wrinkles around these large, trapped water pockets, as observed in ES3 near the Bowtie Section and the Western Section. It is also evident that rainwater has become trapped within the ridges in ES3.

3.5. Displacement Analysis of Floating Cover

The cover-referenced local transformation achieved an alignment error of less than 4.78 pixels (9.6 cm) relative to the predefined feature coordinates. For the displacement analysis, elevation values were adjusted so that the zero value corresponded to the water level. In this study, the observed displacements exceed this alignment bound, so the residual registration uncertainty does not limit the interpretation of relative changes.

Across the seven surveys, NCC-based feature matching identified, on average, 67.4% of features. Identification rates were the highest in July 2023 and November 2023, at 87.0% and 89.6%, respectively. However, identification rates were particularly lower, below 60.8%, in the earlier geospatial data, especially in December 2018, which had the lowest identification rate of 53.3%. The reduced rates reflect temporal changes in natural features that decreased template consistency (e.g., surface appearance changes or occlusion).

3.5.1. Above-Water Elevation of Floating Cover

The maximum above-water elevations show an increasing trend across ES1 to ES4, the Eastern Flotations, and the Bowtie Flotations from December 2018 to November 2023 (see Figure 9, Figure 10 and Figure 11).

Eastern Segments (ES1 to ES4)

• In December 2018, there was a significant maximum elevation increase, with ES4 reaching 46.3 cm and ES1 reaching 67.5 cm.
• There was a rapid elevation increase between December 2018 and September 2019, with ES2, ES3 and ES4 increasing by 17.8 cm to 29.6 cm by September 2019. In contrast, ES1 experienced its rapid increase later, in November 2023, by 11.0 cm.
• The maximum elevations for ES1 to ES4 continued to rise over time, peaking in June 2022, followed by a slight decrease in May 2023, before increasing again in November 2023.
• ES4 consistently showed lower maximum elevations compared to the other segments but experienced the most significant elevation changes: 42.1 cm in November 2023 and 50.1 cm in June 2022 since December 2018.

Top, Centre and Bottom Eastern Flotations

• In December 2018, the top, centre, and bottom Eastern Flotations had maximum elevations of 79.1 cm, 80.7 cm, and 49.7 cm, respectively.
• By June 2022, there was significant vertical movement, with Top, Centre and Bottom Eastern Flotations reaching 118.6 cm, 109.8 cm and 119.6 cm, respectively.
• Bottom Eastern Flotation experienced the largest increase, rising by 13.4 cm to 30.4 cm between December 2018 and June 2022. It also had the most significant overall change of 58.8 cm in November 2023 and 69.9 cm in June 2022 since December 2018.

East Bowtie Flotation

• In December 2018, the initial elevation measurements for the top, centre and bottom of East Bowtie Flotation were 40.4 cm, 45.4 cm and 34.4 cm, respectively.
• The East Bowtie Flotation showed a gradual increase in elevation, particularly between December 2018 and June 2022, where the top flotation reached a high of 68.7 cm in June 2022, representing a 28.3 cm increase since December 2018.
• There was a slight drop in flotation heights in July 2023, particularly for the top and bottom sections, where the elevation decreased to 54.7 cm and 46.0 cm, respectively. Eventually, the flotations increased again by November 2023.

West Bowtie Flotation

• In most instances, the West Bowtie Flotation elevations were lower than those of the East Bowtie Flotation.
• By November 2023, the top, centre, and bottom flotation elevations reached 53.7 cm, 57.5 cm, and 50.6 cm, respectively, showing gradual increases.

3.5.2. Relative In-Plane and Out-of-Plane Movement of Floating Cover

Referring to Figure 12 and Figure 13, the timeline of relative displacement observed in the floating cover relative to December 2018 geospatial data is as follows:

September 2019 (2 years 5 months)

• Eastern Section: Significant increases in relative elevation were observed, with an average increase of 17.7 cm. The Eastern Section showed an average relative elevation increase of 23.6 cm, with a maximum relative elevation of 50.6 cm in ES3 and ES4.
• Bowtie Flotation: The East Bowtie Flotation increased in relative elevation by approximately 20 cm, with the bottom section increasing the most by 35.6 cm. The West Bowtie Flotation increased in relative elevation by approximately 18.0 cm. No significant relative in-plane movement was identified during this period.

November 2020 (3 years 7 months)

• Eastern Section: The maximum relative elevation in ES3 and ES4 increased to 60.9 cm, and the Eastern Section’s average relative elevation rose to 25.3 cm. Additionally, a relative in-plane displacement of approximately 13 cm in the NW direction was observed in ES1 and ES2.
• Bowtie Flotation: A maximum relative displacement of 17.1 cm in the NW direction in the East Bowtie Flotation was observed.

June 2022 (5 years 2 months)

• Eastern Section: A significant increase in relative elevation was observed in ES3 and ES4, reaching 87.7 cm, with the Eastern Section’s average relative elevation increasing to 32.2 cm.
• Bowtie Flotation: The top, centre, and bottom of the East Bowtie Flotation reached maximum relative elevation values of 18.1 cm, 36.7 cm, and 35.0 cm, respectively. The top and centre of the West Bowtie Flotation had relative elevation values of 15.1 cm and 19.9 cm, respectively. Relative in-plane displacements in the NW direction were observed, with the top section of the East Bowtie Flotation showing a maximum relative displacement of 37.3 cm.

May 2023 (6 years 1 month)

• Eastern Section: The maximum relative elevation in ES3 and ES4 was 75.3 cm, and the Eastern Section’s average relative elevation was 27.5 cm. A relative easterly displacement of approximately 18 cm was observed in ES2 and ES3.
• Bowtie Flotation: The top, centre, and bottom of the East Bowtie Flotation reached maximum relative elevation values of 26.5 cm, 37.5 cm, and 37.7 cm, respectively. The top and centre of the West Bowtie Flotation had relative elevation values of 16.5 cm and 24.5 cm, respectively. The East Bowtie Flotation had a maximum relative in-plane displacement of 31.6 cm in the W direction.

July 2023 (6 years 3 months)

• Eastern Section: The maximum relative elevation in ES4 was 68.7 cm, and the Eastern Section’s average relative elevation was 25.5 cm. A relative NW displacement of approximately 20 cm was observed in ES1 and ES3 near the East Bowtie Flotation.
• Bowtie Flotation: The top, centre, and bottom of the East Bowtie Flotation reached maximum relative elevation values of 19.1 cm, 38.5 cm, and 32.7 cm, respectively. The top and centre of the West Bowtie Flotation had relative elevation values of 16.9 cm and 25.4 cm, respectively. The East Bowtie Flotation had a maximum relative in-plane displacement of 31.4 cm in the W direction.

November 2023 (6 years 7 months)

• Eastern and Western Sections: A significant relative elevation increase was noted in ES2, with a maximum of 75.6 cm and the Eastern Section’s average relative elevation increase of 27.2 cm. ES1 near the Bowtie showed a maximum relative in-plane displacement of 37.9 cm in the NW direction. In the lower Western Section, a maximum relative in-plane displacement of 18.0 cm was observed in the S direction. ES1 near the North Boundary showed a localised relative displacement of approximately 20 cm in the E direction, while ES1 and ES2 near the East Boundary showed movement in the S direction.
• Bowtie Flotation: The top and centre of the East Bowtie Flotation relatively displaced 34.6 cm NW and 31.9 cm W, respectively, while the top and centre of the West Bowtie Flotation relatively displaced 27.8 cm and 24.6 cm, respectively, both in the SW direction.

Additionally, localised relative in-plane displacements and elevation changes were observed, which were primarily caused by trapped biogas and rainwater. Figure 9, Figure 12 and Figure 13 illustrate this with an example from July 2023, where a localised elevation change occurred in ES2 near the centre of the East Bowtie Flotation due to biogas pockets. Similarly, in May and July 2023, the presence of trapped rainwater in ES2 and ES3 also led to localised in-plane displacements and elevation changes, as shown in Figure 13. The consistent accumulation of trapped rainwater within the ridges has been observed, with large volumes tending to persist in the same locations over time. On-site inspections have revealed that these ridge areas consist of highly elevated and solidified scum, with the trapped rainwater conforming to these areas. It is also anticipated that fluctuations in the water level, driven by factors such as seasonal rainfall, evaporation, and operational changes at the anaerobic lagoon, cause variation in the boundary elevation of the floating cover.

4. Discussion

The findings indicate that the integrated UAV LP method efficiently produces DEMs and orthomosaics with the resolution required for monitoring a 25W anaerobic lagoon, outperforming those generated by photogrammetry alone. It should be noted that the raw resolution of the geospatial models produced by each method is inherently different and cannot be adjusted due to limitations in both the software and hardware used. Therefore, a direct like-for-like comparison is not possible. However, even at a lower resolution setting, the photogrammetry method requires longer production times than LiDAR at the highest resolution. Therefore, LiDAR-derived DEMs yield superior results. In cases where high resolution is not necessary, the production time for photogrammetry can be significantly reduced. Nonetheless, photogrammetry remains a viable, low-cost, and effective method for generating geospatial data when high resolution and production speed are not critical requirements for the specific application.

In practice, photogrammetric procedures often require multiple iterations to ensure the accurate production of geospatial data, as mentioned previously. This typically involves optimising camera parameters, rebuilding dense clouds, and reprocessing image alignment to reflect updated GCP coordinates. A known issue with photogrammetry-derived DEMs is their susceptibility to geometric distortion, such as offsets and radial distortions [15], particularly in systems using wide-angle (fish-eye) lenses [31], which require additional processing efforts to correct. In our previous UAV systems, which included a bespoke UAV system with an Olympus E-PL7 with a 14–42 mm lens [42], and a DJI M600 Pro with a Zenmuse X5 (15 mm lens) [27], radial distortions were noticeable, and post-processing adjustments were required for correction. This involved additional camera calibration to refine the intrinsic parameters and correct systemic distortion errors, following approaches established in prior studies [15,29,30,31]. Consequently, this process becomes time-consuming due to the repeated validation and calibration steps necessary to correct the geospatial data. In contrast, processing raw LiDAR-derived DEMs is more straightforward, requiring only the inclusion of a single GCP and being free from extreme distortions, thus making them readily usable for analysis.

The integrated UAV LP workflow leverages complementary strengths: LiDAR provides stable absolute elevation, while photogrammetry contributes dense texture for seam-scale interpretation. Consistent with our validation, LiDAR exhibits lower vertical error and photogrammetry yields lower in-plane errors. Accordingly, the fusion anchors elevation to LiDAR and planimetry to photogrammetry. Operationally, LP reduces photogrammetric re-processing by orthorectifying imagery on LiDAR-derived DEMs, avoiding repeated dense-cloud rebuilds and camera re-optimisation. In contrast to LiDAR-only approaches, LP contributes the texture necessary for in-plane displacement mapping and for quantifying wrinkles. In contrast to photogrammetry-only approaches, it constrains geometry to reduce distortion and vertical drift, yielding more consistent results across surveys. Compared with terrestrial surveys (e.g., total station, TLS), the LP workflow delivers full-field, non-contact coverage with shorter mobilisation and requires only a single GCP. These differentiated advantages clarify how the proposed method improves upon existing single-sensor and ground-based practices for SHM of floating covers.

The LP workflow remains sensitive to practical limitations. It relies on the LiDAR-derived DEM as the surface backbone and consequently, local noise and artefacts can propagate into the fused geospatial data. Furthermore, co-registration between LiDAR and optical imagery can be challenging on bespoke or cross-vendor camera rigs, which require accurate intrinsics, calibration, and time synchronisation. In this study an integrated DJI payload was used, which performs on-board alignment and reduces this risk. Additionally, data heterogeneity (i.e., differences in sampling density, incidence geometry) and resolution between LiDAR and photogrammetry can introduce blending and interpolation biases. In this study, the orthorectified images required seamline correction in the Eastern Section near the eastern boundary, where a systematic drift in Easting was observed at the first four porthole CPs (25W-01 to 25W-04). While correcting the orthorectified images was straightforward, ongoing work seeks to determine whether this artefact stems from flight operations, image misalignment, orthomosaic generation, or software-related factors (e.g., proprietary algorithms or optimisation processes).

To mitigate risk, UAV flights were conducted only under benign weather conditions (no rain, minimal wind) and in accordance with WTP protocols. It should be noted that airspace rules also constrained flight altitude, which imposes a trade-off between image GSD and LiDAR incidence angle and can reduce local detail near restricted areas. Nonetheless, the flight survey design followed our prior UAV validation studies [27,42], and operations complied with WTP constraints (e.g., restricted flight parameters) and DJI standard flight limits to maintain accuracy and human and wildlife safety.

While LiDAR and photogrammetry have different error characteristics, CP errors are centimetre-scale for both modalities, and registration in the cover-referenced local frame maintains an alignment bound of 9.6 cm across periods. Given that the measured displacements generally exceed this bound, the residual measurement differences are not significant for cover monitoring or for quantifying relative deformation. Nevertheless, because the workflow was tailored to the WTP floating cover, transferring the technique to other applications may require adaptation.

Furthermore, since the deformation of the floating cover is relatively gradual and not excessively severe, the current image-matching approach using NCC performs effectively. However, in cases of more pronounced or complex deformations, more robust methodologies, such as scale-invariant techniques, will need to be explored to ensure accurate feature localisation. Future work will focus on enhancing and automating feature detection and the monitoring process, with the aim of improving both the speed and efficiency of the analysis.

In this industry research, the availability of highly accurate geospatial data provides deeper insights into scum accumulation in the 25W anaerobic lagoon. The findings presented in this paper reveal increased wrinkling, visually indicating further lateral movements as of November 2023. Additionally, displacement analysis highlights distinct behaviours of the scum. The rapid elevation displacement within the first two years following the installation of the cover suggests accelerated scum growth. It was observed that the top and centre of the Eastern Section (i.e., ES1-ES3) elevated first, followed by the bottom of the Eastern Section (ES4), with similar elevation levels observed by 2022. This was then followed by gradual lateral movement, indicating dynamic forces exerted on the floating cover by the scumberg and the hydrodynamic activity of the continual sewage flow.

The study also uncovers phenomena, particularly in June 2022, when the relative above-water elevation with respect to December 2018 was at its maximum but subsequently decreased, refer to Figure 10 and Figure 11. According to the correlation from our previous study [26], this decrease in elevation suggests a reduction in scum, which is improbable given the expected scum accumulation in the absence of intervention. Another example of this was observed locally in September 2019 in ES2 (see Figure 9), where solidified scum accumulated into a highly concentrated peak, elevating the cover to approximately 94.1 cm, refer to Figure 10a), before decreasing in elevation to approximately 82.8 cm by November 2020. It is anticipated that this correlation may no longer hold as the scum progresses to a later stage, i.e., solidifies and causes significant elevation of the floating cover due to factors such as increased pressure on the cover and high surface water levels. June 2022, in particular, is anticipated to mark the transition to this later stage globally. Therefore, measuring above-water level elevation alone may not be sufficient for assessing scum characteristics during the later stages of scum accumulation in the lagoon.

Nonetheless, these insights into scum dynamics were uncovered through longitudinal assessments and temporal comparisons, facilitated by the availability of high-quality geospatial data on the floating covers. The integrated LP method has been instrumental in augmenting these observations, enabling the effective processing of high-resolution data with precision LiDAR measurements. From an environmental perspective, early detection of structural changes enables timely interventions that can reduce the environmental risk, thereby supporting climate change mitigation and water quality objectives. Ongoing research continues to focus on developing robust monitoring systems and streamlining evaluations of the structural integrity and performance of anaerobic lagoons, with a particular emphasis on scum characterisation and mitigation strategies. These efforts aim to advance the long-term management, environmental sustainability, and operational efficiency of WTP anaerobic lagoons in alignment with Industry 4.0.

5. Conclusions

In this case study, the accuracy and generation of geospatial data for the 25W anaerobic lagoon by leveraging both UAV-based LiDAR and photogrammetry were examined. By integrating both techniques, the strengths of each approach are leveraged to enhance geospatial data development. Specifically, the LiDAR point cloud is used to rapidly generate the DEM, which then serves as the reference data for image orthorectification to construct the orthomosaic, bypassing the bottleneck of photogrammetric depth reconstruction. The results demonstrate that LiDAR-derived DEMs can be produced significantly faster—approximately four times faster—than photogrammetry-derived DEMs, while still achieving the resolution and accuracy required for this specific application. Additionally, the geospatial data obtained through these methods also facilitated the analysis of both in-plane and out-of-plane displacements of the lagoon’s floating cover. The findings indicated that the floating cover elevates rapidly during the first few years after installation, followed by increased lateral displacement. Additionally, the lateral movement is visually manifested through the appearance of wrinkling. This progression corresponds to the behaviour of scum, revealing its growth rate and dynamic impact on the covers. Ongoing efforts in Melbourne Water’s research projects continue to focus on achieving robust structural health and performance monitoring of this critical asset, supporting decision-making and improved strategy development while contributing to long-term environmental sustainability through emission reduction, odour control, and protection of surrounding aquatic ecosystems.