It has been demonstrated that by using SfM [15
] and scale invariant feature transform (SIFT) [16
] photogrammetric-based algorithms, a 3-D sparse point cloud model can be resolved from a series of 60–70% overlapping images [2
]. The method involves the recognition of features that are present in multiple images and calculation of their spatial coordinates by triangulation based on intrinsic (e.g., focal length) and extrinsic (pose and orientation) camera parameters [18
]. A much higher-resolution and dense point cloud model can be reconstructed from the sparse point cloud by multi-view stereopsis (MVS) image-matching algorithms [19
The quality of the image depends on a number of factors, such as camera (sensor size, sensor type, sensor resolution), lens, and camera settings (ISO, aperture, shutter speed). Whereas, the quantity of photogrammetric data, accuracy and aerial coverage are limited by technical factors, such as aircraft maximum flight time and speed, remote control transmission range, camera axis to control vertical, inclined and horizontal orientations of the lens with respect to the Earth’s surface, overlapping ratio, UAV hover effects, and weather conditions. Image sharpness and quality benefit substantially from camera stabilization during flight. Additionally, modern UAVs resolve time synchronization between the camera shutter and on-board GPS and can produce geotagged images.
SfM builds on the principle of generating a 3-D spatial point cloud from a set of unconstrained overlapping two-dimensional (2-D) images that have been taken from different viewpoints. Camera parameters, camera positions, and camera orientations are automatically solved for each image [21
]. Key points that have been matched in overlapping images allow for generating a sparse point cloud, which is only possible for stationary objects. The procedure is generally referred to as the bundle adjustment [21
]. A subsequent dense point could model can be generated by increasing the number of spatial 3-D points while using MVS algorithms, resulting in a high density point cloud compared to the initial point cloud derived from the feature matching process based on SIFT. The model can be georeferenced using GCPs, camera locations (geotagged images), or both [23
PhotoScan software, which is used in this study, is based on SfM-MVS algorithm automatically detecting a suite of tie points in overlapping multiple image pairs, and in turn, recognizes 3-D positions and orientation of the camera, and 3-D location of each feature in the images to generate a dense 3-D point cloud. The dense point cloud forms the basis to produce the orthomosaic, digital elevation model (DEM), and textured mesh. The latter can be imported into any commercial or open source 3-D modelling software packages, such as GEOVIA Surpac (Dassault Systèmes, Vélizy-Villacoublay, France), GOCAD, or CloudCompare for quantitative analysis.
2.2. P3P Aerial Platform
The P3P is a 4-rotor micro-drone quadcopter with a user-friendly remote control, an electronically stabilized digital camera (12 megapixel resolution), GPS, and a maximum flight time of 23 min (Figure 1
). The on-board GPS in the aircraft allows for embedding location information in each image and can be used for direct georeferencing [23
]. An important feature of the P3P is its 3-axis (pitch, roll, yaw) electronic gimbal stabilization technology, which keeps the camera steady and in optimal level during flight or windy conditions (not over 10 m/s), yielding blurred-free images. The gimbal can tilt the camera vertically within 90° range allowing for the operator to collect oblique or nadir imagery (Figure 1
). The transmission distance from the remote control is approximately 2000 m.
P3P comes along with the DJI Go application to remotely capture high-quality still images. The DJI Go application allows for the pilot to see real-time video footage, along with having full aircraft and camera control. Before flying the aircraft, it is obligatory to check the legal flight limitations and requirements from the local authorities. In our case, we followed Trafi regulations of Finland. We recommend that the new users should follow DJI flight video and the safety tutorials available on their website.
2.3. Image Acquisition and GCPs
In this study, images were acquired maually with a camera position orthogonal to the surface with an effective overlap of >80%. The photographic overlap was estimated with the grid lines available on the DJI Go application. Additional oblique images were taken to improve the quality in building the DEMs [25
]. Images were captured during an overcast day for uniform lightening on the surface (Figure 3
). The GPS information and camera parameters were embedded in each image as EXIF (Exchangeable Image File) metadata tags. The EXIF tags provide a real-world coordinate system to sparse- and dense-point clouds in the PhotoScan, resulting in thousands of real-world coordinates that are linked to points in each image, and allowing for the images to be orthorectified.
GCPs provide accurate georeferencing to the 3-D geometry [4
] and can be used to assess the model in a case of direct georeferencing [24
]. A minimum of three GCPs are required, however, depending on the size of the area and accuracy constraints, a uniform distribution of more than three GCPs are preferable [26
]. In this study, ground measurements were performed with a Topcon HiPer®
RTK (Real-Time Kinematic) receiver that allows cm-accuracy along with the FC-200 field collector and TopSURV application (Figure 1
e). The spatial accuracy of the instrument is lower than 10 to 30 mm in all dimensions. GCPs were evenly distributed around the scene. Each GCP is made from a 0.5 × 0.5 m size plastic template and sprayed with a luminous yellow color so that the centre point became always distinct for a precise GCP identification (Figure 1
e). Eleven signals were placed around the mine pit (Figure 3
a), seven were used as control points (Table 1
), and four as check points (Table 2
). Whereas, the known length and width of the dome-shaped warehouse on the western side of the scene is used to further compare the distances with the model (Figure 3