A novel approach to estimate systematic and random error of terrain derived from UAVs: a case study from a post-mining site

In recent years, there has been a major development in the field of Unmanned Aerial Vehicles (UAVs) as well as a significant increase in the use of aerial photogrammetry, which is an affordable alternative to using LiDAR However, the nature of the data obtained from photogrammetry differs from LiDAR data. Photogrammetry using the Structure from Motion (SfM) method is however computationally complicated and results can be effected by many influences. In this paper, data from two UAVs were compared The first one is a commercial eBee system produced by SenseFly equipped with a Sony Cyber-shot DCS-WX220 camera. The other is a home assembled solution consisting of EasyStar II motorised glider and 3DR Pixhawk B autopilot equipped with Nikon Coolpix A camera. The area of spoil heap was measured by both systems in the leaf-off period Both systems were set up identically for data acquisition (overlapping, resolution), which made a comparison of the output quality possible. The ground control points (GCPs) were placed in the study area and their position determined by GNSS (RTK method). A traditional approach for point clouds accuracy validation is their comparison with data of greater accuracy. Unfortunately, the photogrammetry is often validated using GNSS points, the position of which is determined under different conditions than GCPs (different daytime, number, and visibility of satellites, etc.). The magnitude of photogrammetry errors is theoretically the same as that of GNSS. Therefore, in this study, we suggest a novel approach that can be used to compare the accuracy of UAV point clouds without the need for additional validation data (for example, GNSS survey). To exemplify this approach, we used data gathered by two UAV systems (eBee and Easy Star II). Particularly, we statistically estimated the accuracy of the UAV point clouds; used two approaches to estimate standard deviations (one of them using estimated dependencies between data); and investigated the influence of the vegetation cover. To determine the systematic and random errors of the UAV systems data, three areas were selected each with a typical example of vegetation on the spoil heap (forest, grass, bush). A comparison of the individual data in a grassy area suggests that the accuracy of the differences is about 0.03 m, which corresponds to the actual pixel size. Average shift (systematic error) ranged from 0.01 m to 0.08 m. In the forest terrain, the accuracy of data differences is about 0.04 m, which is slightly worse than in the grassy area. Bushy terrain data achieves precision values between a grassy area and a forest area.