Influence of Flight Height on The Accuracy of UAV Derived Digital Elevation Model at Complex Terrain

Lightweight Unmanned Aerial Vehicle (UAV) for 3D topographic mapping in mining industry has been raised significantly in recent years. Especially, in complex terrains such as in open-pit mines in which the elevation is rapidly undulating, UAV-based mapping has proven its economical efficiency and higher safety compared to the conventional methods. However, one of the most important factors in UAV mapping of complex terrain is the flight altitude, which needs to be considered seriously because of the safety and accuracy of generated DEMs. This paper aims to evaluate the influence of the flight height on the accuracy of DEMs generated in open-pit mines. To this end, the study area is selected in a quarry with a complex terrain, which is located in northern Vietnam. The investigation was conducted with five flight heights of 50 m, 100 m, 150 m, 200 m, and 250 m. To assess the accuracy of resulting DEMs, ten ground control points (GCPs), and 385 checkpoints measured by both GNSS/RTK and total station methods were used. The accuracy of DEM was assessed by root-mean-square error (RMSE) in X, Y, Z, XY, and XYZ components. The results show that DEM models generated at the flight heights of less than 150 m have high accuracy. RMSEs of the 10 GCPs increase from 1.8 cm to 6.2 cm for the vertical (Z), and from 2.6 cm to 6.3 cm for the horizontal (XY), whereas RMSE of 385 checkpoints increase gradually from 0.05 m to 0.15 m for the vertical (Z) when the flight height increases from 50 m to 250 m.


Introduction
Digital Elevation Model (DEM) or Digital Terrain Model (DTM) is a three-dimension (3D) representation of a bare-earth model (Guth, 2006), while Digital Surface Model (DSM) is a 3D representation of the Earth's surface, which includes all the canopies, buildings and other man-made features (Li, Zhu, & Gold, 2004). Both of the two models have their own specific applications (Fleming, Marsh, & Giles, 2010;Růžičková & Inspektor, 2015), but in open-pit mines, due to the absence of vegetation coverage or man-made features, DEM and DSM can be safely treated interchangeably as is done in this manuscript hereafter.
In open-pit mines, a precise DEM is necessary that is used for various applications, e.g., stockpile management, pit and dump management, slope stability, and mining-induce subsidence ( (Farr et al., 2007). Although these DEMs have been used widely in various fields, e.g., hydrology, geophysics, environment, they are ineffectual in open-pit mines where the landscape changes abruptly within a few meters due to their coarse resolution (usually at 30 m or 90 m). Additionally, they are not updated frequently, and thus do not represent the most recent Earth's surface in mining areas, which is changing continuously during a mining project.
Local DEMs are therefore frequently employed in openpit mines rather than a freely available global DEM. They are usually generated using either conventional surveying by, e.g., automatic levels or total stations, or more modern methods, which includes the use of 3D point cloud from LiDAR or UAVs photogrammetry. The conventional methods provide a high precision but have a limitation of being labor-intensive with high cost and time. Hence, it is not a feasible method to generate an enormous number of DEMs required during the course of open-pit mining projects. Alternatively, LiDAR is an advanced technique for the rapid collection of the 3D point data of the topography, which is used to create reliable DEMs. However, the cost of collecting and processing the LiDAR data is very high, which may in turn increase the total cost of a project significantly. Conversely, UAV photogrammetry has been proven to be a suitable candidate technique for DEM generation in open-pit mines, which is neither comparatively labor intensive nor costly. . The safety and productivity of the UAV-based method are dependent on several factors, in which the flight altitude is one of the most important factors. This is due to the fact that the flight altitude has a direct impact on the resolution of UAV images in such a way that the higher the flight altitude is, the lower resolution the generated DEM will have and vice-versa. This further affects the accuracy of the produced DEM.
From the safety perspective, the flight altitude needs to be considered seriously because the maximum flight altitude limitation is set by local authorities, whilst the minimum flight altitude depends on other factors, such as the accuracy requirement and the UAV take-off position. In terms of productivity, the amount of data captured and the processing time are inversely proportional to the flight height, and these factors may broadly correlate with the DEM accuracy as well. Therefore, a detailed investigation of the influence of the UAV flight height on the DEM accuracy is indispensable, especially when the UAV technique has been advocated as a better choice than others for efficient generation of multiple DEMs in a mining project.
There have been several studies on the influence of the UAV flight height on the accuracy of DEM (e.g., (Fuad et  In the next section, the study area and methodology are discussed, which are followed by the results and discussions shown in Section 3. The conclusions of the study are shown in Section 4 with future recommendations.

Study area and data collection Study area:
This study is conducted in the mining site of Long Son limestone quarry that belongs to the Long Son cement plant. This quarry is located in Ha Vinh commune, Ha Trung district, Thanh Hoa province, which is about 200 km away from Hanoi, Vietnam. It encloses an area of about 1 km2 with topographical characteristics of a quarry similar to many other quarries in northern Vietnam, including benches, the toe of the benches, and steep slopes. At the time of the study, the mine had already gone through the stage of basic construction and was at an excavating level of 110 m (Fig. 1).

Survey flight:
To collect aerial data, a commercial quadcopter DJI Phantom 4 Pro mounting a 20-megapixel RGB camera with a focal length of 8.8 mm and a sensor size of 13.2 mm x 8.8 mm (Https://www.dxomark.com) that allows high-resolution aerial photography is employed. Its airframe carries a GPS/ IMU that enables it to have a posture control, stop flight, and automatically take off and land with high stability. The drone operates in both manual flight mode using a controller and automatic flight mode using an Android or IOS smartphone. In this study, we use Pix4Dcapture software installed on an iPhone 7 plus for planning the flight. In the automatic mode, several vital parameters are set to the drone, including a mapping area of 1 km2, a flight height, forward overlap, and side-lap of images of 80%. The position of each camera was measured at the time of shooting by a low-cost GNSS/INS receiver mounted on the drone, with an average accuracy of 2.5 m along the axes. This information is stored in each image's file and used for processing imagery to obtain photogrammetric products.
To assess the influence of flight heights on the accuracy of the DEM, we flew the Phantom 4 Pro at five different altitudes above ground level: 50 m, 100 m, 150 m, 200 m and 250 m with average Ground Sample Distance (GSD) of 1.36, 2.73, 4.09, 5.45, 6.82 cm/pix, respectively. The take-off location of the UAV is chosen to ensure safety, especially when flying at low altitudes. In this study, we choose the take-off location located at the middle high of the study area, with an elevation of 80 m above sea level, which ensure that when we fly the UAV at the lowest altitude of 50 m, the UAV is safe from the highest point in the study area (Fig. 2b).

Establishưding the Ground Control Points:
Ground Control Points (GCPs) play an essential role in geo-referencing and evaluating the accuracy of DSMs. Therefore, GCPs need to be placed on the ground before carrying out image acquisition. In a condition that quarry were still operating, field reconnaissance with handheld GPS is conducted to select safe areas for placing these GCPs. A total of eight GCPs were used for the calibration (camera-lens optimization and bundle block adjustment) and generation of the different DEMs, while ten GCPs are used as checkpoints (Fig. 3). The 18 GCPs were placed at various heights and uniformly distributed to cover the entire area. These were measured by dual-frequency GNSS/RTK receivers (Fig. 4). The specifications of GNSS receivers are listed in Table 1. In addition, to better assess the accuracy of DSMs, more than 300 points, which are measured by the same method over the area on main fea- tures including benches, the toe of the benches, and around 80 points measured by the total station with non-prism for safety over the area on slopes as these areas are complex terrains and cannot be accessed in person, the expected accuracy in elevation of these measurements was to be about 1-1.5 cm. A Leica TS02 total station with an angular accuracy of 5" and a distance accuracy of 1.5 mm + 2 ppm was used (Fig. 4). The GCPs and checked points are marked on the ground with the size of 60x60 cm targets which were designed in black and yellow with a highly reflective material. The coordinates of GCPs were measured in the national control network (the VN2000 coordinate system), using the GNSS/RTK method. The base receiver was installed at one control point of the national control network established in the local area; the rover measured each point with a pole clip to ensure stability. Each point was measured at least three times, and after eliminating rough errors, the final point coordinates were the average value. Any GCPs that did not satisfy the accuracy were removed, and the remaining points were used to calibrate the camera and evaluate the accuracy of models.

Software and image processing:
In this study, Agisoft Photoscan Professional (ver 1.54) was used to process UAV images acquired in the field. The data processing procedure of Agisoft Photoscan includes five steps: (i) photo alignment; (ii) bundle block adjustment; (iii) optimization, (iv) 3D surface reconstruction, (v) generation of DSM. Firstly, when images are input, key points that can be identified from each image are automatically extracted, and then the singularities extracted from each image are linked with each other through mutual comparison among a plurality of images. This process is called "photo alignment". When geometric corrections of singularities are completed in each image, the singularities representing the same points in several images are automatically matched through a comparison between consecutive UAV images. When the singularities are extracted, the 3D point cloud and the DSM model are generated.

Evaluation Methods:
The goal of this study was to evaluate the accuracy of DEM generated by UAV with different flight altitudes. For this, the block orientation quality was first assessed on 10 signalized 3D checkpoints. Further, the DEMs were evaluated by comparing the observed and extracted elevation values at a broader set of checkpoints using the different values and the Root Mean Square Error (RMSE) given by Eqs. 1-6. where X GCPi and X DSM are the X-coordinate component of GCP and corresponding coordinate in DSM, respectively; Y GCPi and Y DSM are the Y-coordinate component of GCP and corresponding coordinate in DSM, respectively; Z GCPi and Z DSM are the Z-coordinate component of GCP and corresponding coordinate in DSM, respectively.

Results and Discussions
The accuracy assessments of the camera-lens model calibration and the DEM with eight GCPs and 10 checkpoints for 50 m flight height are summarized in Tables 2 and 3, respectively. Figure 6 depicts the mean ground accuracy measured (RMSEX, RMSEY, RMSEXY, RMSEZ) with respect to the five different flight heights.
At the flight altitude of 50 m, the largest number of images (247) at a resolution of 1.36 cm/pixel were acquired. The point cloud of the quarry with approximately 101,982 3D points was extracted following the method described above. Further, an orthographic image and a DEM were generated with resolutions of 1.3 cm and 13.3 cm/pix, respectively. Table 2 shows that the accuracy of the model built from the calibrating dataset is much reliable, with all the errors within 1 cm. However, it may be too positive as the calibrating dataset was used for both the optimization process and the goodness-of-fit. Therefore, the checking dataset was used to assess the accuracy of the DEM model. In Table 3, it could be seen that the maximum error for X is -3.9 cm (GCP9), Y is 3.4 cm (GCP19), Z is 3.1 cm (GCP17), XY is 3.9 cm (GCP9), and XYZ is 4.7 cm (GCP17). With a DEM of 13.3 cm resolution, RMSEXY of 2.6 cm and RMSEZ of 1.8 cm indicates the appropriate reliability of the constructed DEM to be used in open-pit mining applications.
From Table 4 and Figure 5a it is observed that the RMSE in X, Y, XY, Z, XYZ is smallest with 0.  (Table 4). Figure 5b shows that for the checking dataset also the smallest and largest RMSE in X, Y, XY, Z and XYZ are with the flight altitudes of 50 m and 250 m, respectively. Thus, we obtain an indirect proportionality between the flight altitude and the DEM accuracy. This is further confirmed by comparing each DEM with the elevation model constructed using 385 GNSS and total station points measured on benches, toe of the benches and steep slopes. Figure 5

Conclusions and Proposals
An experimental investigation on the accuracy of DEMs generated from UAVs was performed by flying the drone at five different altitudes 50 m, 100 m, 150 m, 200 m and 250 m above ground level over complex terrains of open pit mine. Accordingly, a lightweight and the 1-inch EXMOR R CMOS camera mounted on a low-cost DJI Phantom 4 Professional UAV was used. The acquired number of images increased from 80 to 247, with a decrease in the flight altitude from 250 m to 50 m. The processing and accuracy assessment were carried out using 18 GCPs and 385 checkpoints.
The result showed that the accuracy of DEM is reliable, with flight height ranging between 50 m to 150 m. RMSE in the calibrating dataset increases from 0.5 cm to 4.4 cm for vertical, and from 0.4 cm to 3.4 cm for horizontal, indicating high success-rate of fit in processing, whereas RMSE in the checking dataset increases from 1.8 cm to 6.2 cm for vertical, and from 2.6 cm to 6.3 cm for horizontal, indicating high accuracy. These indicate that the processes of capturing images, the establishment of GCPs, and photogrammetric processing were carried out successfully. Also, it is concluded that the altitude of the drone influences both horizontal and vertical accuracy of DEM. The greater the flight height is, the lower the accuracy of the DEM model.