Supplementary material to "A Structure from Motion photogrammetry-based method to generate sub-millimetre resolution Digital Elevation Models for investigating rock breakdown features"

Abstract

Stage 1: All images are loaded and reviewed in Photoscan. Data with image quality value higher than 0.5 were selected (Agisoft, 2016). The coded GCPs were detected automatically before starting batch processing of images.Local coordinates for each GCP target was entered into Photoscan. The measurement accuracy was adjusted to 0.01 mm for marker accuracy and scale bar accuracy in reference settings for experiments. This value was 0.5 mm for control target in the field. Unwanted scenes in the background or foreground of images (e.g. sky and plants) can lead to incorrect point matching. Photoscan offers a solution to mask out unwanted areas. While this can be a time-consuming process, it can be minimised with careful image acquisition.Stage 2: Images alignment and optimisation: The program matches the common point in images and determines the camera position for each photo. A 3D sparse point cloud is generated using a least square solution (Thoeni et al., 2014; Agisoft, 2016). The error between measured GCP coordinates and estimated GCP coordinates is determined through the least squares solution (Thoeni et al., 2014). Photos were aligned with the highest alignment accuracy, generic pair pre-selection and the default per-image key and tie point limits. To generate the most accurate 3D model, it is crucial to overcome systematic dishing and doming distortions in the SfM model by correcting lens distortions (Carbonneau and Dietrich, 2016). SfM workflow performs a self-calibration using Exchangeable Image File Format (EXIF) metadata from digital images. Each image is treated as unique during the self-calibration process. The GCP errors were minimised using the camera calibration parameters to refine any distortion in the model during optimisation. PhotoScan adjusts estimated point coordinates and camera parameters, during optimisation, thus reducing the sum of reprojection error and reference coordinate misalignment error (Agisoft, 2016). Although a 3D sparse point cloud is not required for DEM generation, it is required for dense point cloud reconstruction.Stage 3: Dense Point cloud generation: The dense point cloud is built, using estimated camera positions from sparse point cloud generated during image alignment process. A range of quality options are available, and we selected "High". This decision was based on the time required to achieve the required quality for our work.Choosing "Ultra high" can result in higher point density but increases processing time. We used "Mild" depth filtering as we wanted to reconstruct smaller breakdown features (Agisoft, 2016). The dense point cloud was used to generate the DEM.