Multi-view stereopsis (MVS) from an unmanned aerial vehicle (UAV) for natural landform mapping

Unmanned aerial vehicles (UAVs) or drones have become a cost-effective tool for surveying and mapping. UAV photogrammetry using Multi-View Stereopsis (MVS), known as UAV-MVS, is a technique that combines photogrammetry and computer vision and is becoming increasingly popular for 3D reconstruction surveys. There is a need for rigorous accuracy assessment of 3D reconstructions using UAV-MVS. This thesis evaluates UAV survey design considerations (camera network, camera calibration, and ground control network) and assesses their impact on the accuracy of MVS point clouds. The aim of this thesis is to assess the accuracy of the UAV-MVS survey technique to better understand the scale of change that can be detected. The chosen application area is natural landform change, in this case of a section of sheltered coastline that is eroding at scales that are difficult to monitor from satellite or aerial photography. Quantifying the spatial and temporal scales of the erosion occurring along these often fragile sheltered coasts provides an insight into the response of the landscape to many different variables including sea level rise. The erosion is occurring at the decimetre and centimetre scale and UAV-MVS from low altitude (<40‚Äö-50 m) can provide 3D point clouds with 1‚Äö-6 points per `cm^2`. These point clouds are compared to an in situ verification dataset derived using a total station survey (˜ìvâ ranging between 1 and 2 mm) to assess accuracy. Various UAV survey design and processing scenarios are compared to assess the impact of camera calibration method, image overlap, inclusion of oblique imagery, ground control survey precision and ground control point (GCP) distribution and density. The use of simulations to predict achievable accuracy is tested. Profiles are used to compare cross sections of the point clouds and visualise point density, accuracy and detected change. Dense point clouds are commonly converted to digital elevation models (DEMs) and differenced to detect and quantify change. The conversion process can introduce artefacts into the data through the interpolation and generalisation process. For this reason the change detection and quantification methods evaluated here are based on point cloud differencing and extracted shoreline comparison. Vegetation edge and erosion scarp edge are two shoreline proxies that are extracted and compared. The results demonstrate that UAV-MVS point clouds of natural terrain can be accurate to < 5‚Äö-6 mm when using precise control (˜ìvâ = 1‚Äö-2 mm) and 10‚Äö-11 mm when using differential GPS equivalent control (˜ìvâ = 22 mm). The flying height in these tests was ~20‚Äö-25 m above terrain. Comparisons of network simulations and empirical data demonstrate that a simulation can be used to reliably predict object space accuracy for typical UAV operations, where ground control is being established using differential GPS (˜ìvâ = 22 mm). However, simulation predictions are less reliable when the ground control is established using precise field survey methods (˜ìvâ = 1‚Äö-2 mm), with achieved accuracy lower than predicted precision. This is attributed to the influence of residual systematic errors in camera calibration. The findings further demonstrate that 70‚Äö-80% overlap nadir photography supplemented with oblique photography focussed on complex portions of the terrain provides the most accurate and complete 3D reconstructions of this coastal shoreline. Point cloud differencing is an effective means of detecting and quantifying change. When extracting a proxy for shoreline at the fine-scale provided by UAV-MVS data, the scarp edge is more easily delineated than vegetation edge and provides a more accurate indication of shoreline position and associated change. The use of UAVs for surveying and monitoring can now be undertaken with confidence provided the design guidelines offered in this thesis are adhered to. The use of UAV-MVS to monitor centimetrescale change along sheltered coastlines can provide the spatial and temporal resolution datasets needed to distinguish event-driven change from longer term trends.