Introduction to the Thesis Project
Marc Baauw's research was driven by the aim to gain valuable insights into the differences between two approaches for rugosity calculation using a Structure-from-Motion (SFM) based photogrammetric workflow. The SFM process presented certain challenges, with the usage of low-quality imagery and insufficient overlap proving to be limiting factors. However, through perseverance and dedication, high-quality data was obtained, enabling the generation of a reliable 3D model using the proposed workflow.
A significant breakthrough in Baauw's research was the successful automation of basic SFM processes, which provided valuable insights into the necessity of manual intervention or supervision in quality improvement and scaling procedures. In particular, Baauw discovered the crucial importance of scaling and orientation of the model, as these factors are essential for performing reliable measurements. Despite relying on visual estimations, scaling yielded reliable outputs, contributing to the overall success of the project.
Baauw conducted this research as part of his MSc studies under GIMA (Geographical management of information systems), a combined program offered by TU Delft, Wageningen University, Utrecht University, and the University of Twente. His meticulous efforts and exceptional performance throughout the project are commendable, showcasing his passion and dedication to advancing coral reef conservation. Under the guidance and supervision of Ned Dwyer from Reef Support and Harm Bartholomeus from Wageningen University and Research (WUR), and data contributions from REEFolution Foundation in Kenya and Curacao (Van Zeil, 2021), Baauw's thesis project benefited from the expertise and support of experienced professionals.
Structure-from-Motion (SFM) and the Digital Elevation Model (DEM)
In addition to the primary research focus, an automated photogrammetry workflow emerged as a topic of great interest for the internship organization. This prompted a thorough investigation into the potential automation of processing steps within Agisoft Metashape. Custom Python scripts were developed, leveraging Agisoft's Python API reference, to assess the feasibility of automation.
The research outcomes provided valuable insights into the overall structure-from-motion (SFM) workflow in Agisoft Metashape and the implications of automation. It was discovered that while a partially automated workflow could be achieved, a fully automated approach might not be suitable for every project. By following the basic workflow steps, such as project saving, photo addition, photo alignment, camera optimization, dense cloud generation, mesh creation, texture generation, and output export, visually appealing textured meshes could be generated using the custom scripts.
However, it became evident that certain quality improvement steps, such as gradual selection and confidence-based filtering, required manual supervision or intervention and were not amenable to full automation. Scaling, a critical aspect for specific processing steps like building a digital elevation model (DEM), was deemed essential for utilizing SFM outputs in subsequent calculations. The research highlighted that manual intervention for scaling the model was best performed after photo alignment.
The research findings emphasized that a fully automated workflow would be desirable only when meeting the basic project requirements. However, for projects involving scaling and additional quality improvements, a semi-automated workflow involving manual intervention was recommended. This approach ensured better control and accuracy in achieving the desired outcomes, striking a balance between automation and manual involvement.Throughout the project, several significant milestones were accomplished. Firstly, all 687 images were aligned using the initial settings, resulting in a sparse point cloud comprising 383,442 points. The alignment process took approximately 46 minutes and 56 seconds, while optimization was completed in less than a minute. Manual scaling, including placing, checking, and correcting markers in each image, required around 1 hour. The scalebars associated with the markers demonstrated errors of 0.0022373 and -0.002226 meters, respectively, well below the defined maximum error threshold of 0.005 meters. Figure 12 provides a visual representation of the resulting sparse point cloud after photo alignment.
Furthermore, a dense cloud was generated, consisting of approximately 11,408,646 points, processed at a "medium" quality level. The entire process took 1 hour and 22 seconds. Subsequently, mesh creation based on the dense cloud required 27 minutes and 49 seconds, resulting in a mesh with 3,155,394 faces and 1,578,565 vertices. The texturization process took 8 minutes and 48 seconds. Due to processing difficulties at the initially planned texture size of 16,384, the texture size had to be downscaled to 8192, resulting in a slightly less detailed texture map. Figures 13 and 14 showcase the resulting dense point cloud and textured mesh, respectively.
Another important outcome was the generation of a digital elevation model (DEM). The DEM was created using a planar projection (Top XY) in a local coordinate system, with a resolution of 0.00166334 (1.66 mm/pixel). The generation process took 18 seconds. Figure 15 depicts the resulting DEM, visualized as a shaded relief map.These milestones signify significant progress made throughout the project, showcasing successful generation of both sparse and dense point clouds, as well as the creation of a detailed mesh with accompanying textures. Each step, including alignment, scaling, mesh generation, and DEM creation, was executed with precision, resulting in accurate and visually appealing outputs for further analysis and assessment.
Rugosity Analysis
To calculate both linear and surface rugosity values, Baauw utilized ArcGIS Pro, following established methodologies described in previous studies (Burns et al., 2015; Fukunaga et al., 2019; Chen & Dai, 2021; Pascoe et al., 2021). Baauw created a new feature class and generated a 2 by 2 meter area by creating a rectangular buffer around a central point on the Digital Elevation Model (DEM). This rectangular area was further divided into a grid of 25 equally-sized rectangular areas, each measuring 0.16 m2 (0.4 by 0.4 meters). These grid cells served as separate areas for calculating surface rugosity, and the process is illustrated in Figure 10.
For calculating linear rugosity, Baauw created a grid of 10 by 10 cells. The lines separating the rows and columns of these grid cells were used to determine the linear rugosity values. Specifically, the grid lines extended from one edge to the opposite edge of the 2 by 2 meter rectangle, resulting in 2-meter-long lines. To facilitate direct comparison between the two rugosity values for each grid cell, line segments of 0.4 meters were created by merging the middle cross-sections of each grid cell.
This merging process formed combined line segments measuring 0.8 meters, as depicted in Figure 11.These calculations and preparations in ArcGIS Pro allowed Baauw to determine both surface and linear rugosity values, providing valuable insights into the structural complexity of the coral reef habitat.
In the calculation process described, rugosity values were determined using specific formulas and tools. To calculate linear rugosity (LR), the 'Add Surface Information' tool was employed in ArcGIS, considering both the 2D line length and the 3D surface length derived from the Digital Elevation Model (DEM) as the surface layer. The formula used for LR calculation was as follows:
LR = Surface distance / line length
For surface rugosity (SR), the 'Add Surface Information' tool and the DEM were again utilized to calculate the 3D surface area for each grid cell. The planar area of each cell was known to be 0.4m2. Using this information, the SR values were determined using the following formula:
SR = 3D surface area / 2D planar area
By applying these formulas and leveraging the 'Add Surface Information' tool, Marc Baauw was able to calculate both linear and surface rugosity values. These calculations provided crucial insights into the structural complexity and variability within the coral reef ecosystem.
The Digital Elevation Model (DEM) generated in Agisoft Metashape was exported and imported into ArcGIS Pro. A grid of 25 rectangles, each measuring 0.4 m², was created using the proposed workflow. Two sets of lines were generated: combined lines consisting of two 0.4 m line segments crossing in the middle of each grid cell, and straight lines running horizontally and vertically across the area.
Statistical analysis was performed on the values obtained for surface rugosity and the two linear rugosity approaches. The Shapiro-Will test indicated that these values were normally distributed, as evidenced by their P-values (0.08, 0.30, and 0.31) exceeding the significance threshold of 0.05. The distribution and normality of the data were further visualized through distribution plots and Q-Q plots. While slight skewness was observed in each approach, no significant outliers were present.
A Pearson's correlation test revealed a relatively high positive correlation between surface rugosity and linear rugosity, with an 'r'-value of 0.803. This correlation was statistically significant (P < 0.001), indicating a strong relationship between the two variables. The scatter plot displayed in Figure 20 visually demonstrates this linear relationship between surface rugosity and linear rugosity.
These findings confirm the association between surface and linear rugosity measures and provide statistical evidence supporting their correlation.
Conclusions
The aim of this research was to compare two approaches for calculating rugosity using a structure-from-motion (SFM) photogrammetric workflow. Although there were some challenges with the SFM process, the proposed workflow successfully generated a reliable 3D model using high-quality data. Attempts to automate basic SFM processes were also successful, providing insights into the need for manual intervention in quality improvement and scaling. Scaling and orientation were found to be crucial for accurate measurements, and the output obtained through visual estimations proved reliable.
However, the comparison between the two approaches was hindered by the lack of independent validation data, preventing quantitative assessment and determination of which method better represents structural complexity. Nevertheless, a qualitative assessment compared to similar research revealed that linear rugosity values aligned with expectations, while surface rugosity values were generally higher. The high density of measurements in a small area of the coral reef led to significant relative differences. It's important to note that the values presented in this research may not be representative of the broader environment without multiple samples.
Implementing both approaches in the same area resulted in higher average and dispersion values for surface rugosity compared to linear rugosity. This discrepancy arises from the different dimensions and calculations used by each approach, indicating that surface rugosity and linear rugosity cannot be used interchangeably, despite their usage in contemporary research.
Overall, despite its limitations, this research provides valuable insights into photogrammetric processes and rugosity calculations, serving as a preliminary guideline for future research interns interested in this topic. For full workflow and additional information on data collection, 3D photogrammetry creation and analysis, please contact us.
Conduct a research project with us
This project is a project very close to our hearts at Reef Support. It isn’t only about helping researchers work with coastal communities in their effort to innovate with digital technologies and monitor the reefs but also about fostering wide-spread marine knowledge. For interested students that want to take on a future challenge in reef-related research, do contact us and let us know about your idea, or if we have an opportunity we will see where you can play a role.
