IMPACT OF FLIGHT ALTITUDE ON UNMANNED AERIAL PHOTOGRAMMETRIC SURVEY OF THE SNOW HEIGHT ON MOUNT LEBANON

In Lebanon, the seasonal snowpack is poorly monitored despite its importance for water resource supply. The snow accumulates on Mount Lebanon in karstic depressions named “sinkholes.” It is important to monitor the evolution of the snow height inside those “sinkholes”, because of their key role as “containers” for seasonal snow. UAV photogrammetry is a major technological breakthrough which allows an accurate monitoring of the snow height. Because the impact of flight parameters on snow height retrievals is not well documented yet, this research aims to evaluate the impact of UAV flight altitude on the resolution and accuracy of the resulting orthomosaic and DSM. The flight missions were done using the Phantom DJI which generated five DSMs. These are validated using total station measurements. The results indicate that the snow DSMs can be retrieved by adopting a resolution of 8 to 84 cm, a point density between 1.43 and 153 points/sqm and a RMSE of 13 to 41 cm. The testing was done using an elevation varying between 50 and 500m. The results will be compared to total station observations. These results allow the user to choose the suitable flight altitude for required resolution and points density. We suggest that a flight altitude of 100 m is sufficient for the survey of the snow cover elevation.


INTRODUCTION
In many Mediterranean regions, meltwater runoff from the seasonal snow cover provides essential water resources (Fayad et al., 2017). This is the case in Lebanon where most of the precipitations fall as snow in winter in high elevation regions (Shaban, 2011;Shaban et al., 2004;Telesca et al., 2014). The snow covers around 25% of the total Lebanese territory in winter (Shaban et al., 2004;Telesca et al., 2014;Mhawej et al., 2014). It is estimated that at least 30% of rivers and spring discharges originate from the snow melt (Telesca et al., 2014;Mhawej et al., 2014;Fayad et al., 2017). Although Lebanon is one of the richest countries in terms of precipitations in the Middle East, water restrictions are common during summer. This is partly due to the mismanagement of available water resources, but also to climate change which has caused severe droughts in recent years. A continuous and accurate monitoring of the snowpack is of utmost importance to improve the management of the lebanese water resource and to understand the sensitivity of the snow resource to climate (Somma et al., 2014;Fayad, 2017). Numerous karstic depressions, "sinkholes", punctuate Mount Lebanon high plateaus and play the role of "containers" for the seasonal snow (Somma et al., 2014). They slow down the melting process as they protect the snowpack from the wind and sun radiation. This process delays the low flow period (Shaban, 2011;Somma et al., 2014). Therefore, it is useful to focus the observation of the snow cover evolution inside those depressions in particular to anticipate the beginning of low flow season.
About 50 years ago, pioneering studies were devoted to map snow height by photogrammetry using scanned aerial imagery (Smith et al., 1967). Cline (1993) studied the topic in detail (1993,1994) with mixed results. Recently, the development of Unmanned Aerial Vehicles (UAVs) equipped with a digital camera for Digital Surface Model (DSM) generation has opened new perspectives for the study of the snow cover. The snow height distribution can be obtained from UAV surveys at submeter resolution with decimetric to centimetric vertical accuracy (Bühler et al., 2016;Harder et al., 2016;Marti et al., 2016, Cimoli et al., 2017Avanzi et al., 2018;Lendzioch et al. 2019). The snow height is obtained by subtracting a snow-on DSM and a snow-off DSM, computed with a structure-frommotion (SfM) algorithm of overlapping photographs taken by the digital camera onboard the UAV. In addition to the good accuracy of UAV-derived snow height maps, UAV are increasingly used in snow science as well as in other fields of geoscience (Kelleher et al., 2018;Niedzielski 2018) because they can be operated without a licensed pilot onboard and provide a low-cost solution in comparison with manned aircraft surveys. Not only are UAVs safer to use than an aircraft but they can also fly very close (few meters and above) to the ground (Brunier et al., 2016). However, UAVs offer a limited spatial coverage compared to a manned aircraft (Nolan et al., 2015) or even satellite stereoscopic images (Marti et al. 2016;Deschamps-Berger et al. 2020;Shaw et al. 2020).
A key parameter to increase spatial coverage during a UAV survey is the flight altitude. Higher flight altitudes make it possible to cover larger areas over the same time period , but at the cost of a coarser spatial resolution and accuracy in the final DSM (Leitao et al., 2016;Doumit, 2018;Doumit and Pogorelov 2018).
To our knowledge, this tradeoff has not been documented in the context of snow-covered areas. Here, we apply UAV photogrammetry and SfM to compute DSMs by testing the speed, altitude flight, and processing on different resolutions (low, medium, high). All tests were validated by total station instrument measurements.

STUDY AREA
This study is focusing on an area located in Jabal El Dib in the upper mountains of Ouyoun al-Siman at 2,300 meters above sea level ( Figure 1). The studied site has an area of around 13500 sqm with an altitude between 2313 m to 2330 m above sea level. It represents a pilot site for snow studies in Lebanon (Mzaar weather station, Fayad et al. 2017) as the snow can remain until the end of July or August.

Field measurements
UAV photogrammetric surveys were done using a "Phantom 3 Advanced DJI" (Figure 2). This UAV is equipped with a builtin camera of 12 megapixels, and with an autonomous 17minutes of flying for each fully charged battery. The Positions of the cameras in the global coordinate system were measured during the flight using a standard GNSS receiver mounted on the UAV, without using the real-time kinematic (RTK) solution. In this context the camera position and orientation were estimated by using ground control points (GCPs) during the photogrammetric restitution process (Aguera-Vega et al., 2016).
Total stations ( Figure 3) have been until now one of the main tools used for millimetric precision and accuracy in surveys field, including topography, hydrography, construction, geography, hydrology (Lavine et al., 2003). In this study, the "Leica TS06" total station was used to (i) measure the nine GCPs ( Figure 4) (ii) measure 187 points on the snow cover DSM (Figure 4) as a reference dataset to validate the DSM obtained by UAV photogrammetry.
The points dataset measured with the total station is considered as ground truth with maximum errors of 1 cm and one point, each 5 to 10 m ( Figure 4) . Although, this instrument does not provide high density in comparison to a Terrestrial Laser Scanner (TLS) and Structure from Motion (SfM) that could reach much more than 200 points/sqm; we can be sure of the accuracy of each measured point individually. The accuracy of the UAV DSMs was estimated using the residuals of the elevation between the total station measurements and the DSM value at the point location. Standard deviation, means and maximum errors are used to evaluate the accuracy of each extracted DSM from the different flight altitudes.   In order to respect aerial photogrammetric rules, suitable flight planning should be designed and carried out (Ruiz et al., 2013;Gandor et al., 2015). Aerial photos without enough overlap become useless for generating DSMs and orthophotos. Accordingly, mission planning is considered as a critical phase for successful DSM generation (Cimoli et al., 2017). Five missions were prepared, each one having a different altitude flight ( Figure A1). The flight mission parameters were set using a mobile autopilot Android application named Litchi. As for the interval between two consecutive images, it was adjusted to two seconds which was the minimum interval choice. The five UAV missions were conducted with the same overlap (70-80% along the flight line and 60-70% as side overlap ( Figure 5) following the guidelines proposed by Agisoft (2016). The Litchi application allows the placing of waypoints to have parallel lines, which will represent the autopilot path of the drone while carrying out the mission.
Many times this experiment was unsuccessful due to technical and logistical problems such as low battery, wind speed or lack of authorization. However, on March 28th 2017 we achieved the five flights as planned FA=50m, FA=100m, FA=200m, FA=350m and FA=500m, .

Data processing
The processing was performed on a custom-built laptop running an Intel(R) Core (™) i7-6700 HQ CPU @2.60GHz 2.59 GHz with 4 cores and 32 GB of memory DDR4. The computer GPU is an Intel® HD Graphics 530 DDR4. All the processing of aerial images was done using Agisoft Photoscan. Agisoft allows to choose a dense cloud quality and the level of alignment accuracy (Morgan et al., 2016). For the latter we tried to adjust and change the alignment accuracy setting from "lowest" to "medium" and then "highest". We then alternated the dense cloud quality setting between "lowest","medium" and "ultra-high".
As shown by (Abou Chakra 2018) regarding the quality of processing under Agisoft Photoscan, the "medium resolution" processing has been chosen. We also tested the impact of flight speed by performing two flights; one with the maximum speed of 54 km/h and another with a speed of about 13 km/h. The results suggest that the speed change did not modify the DSM ( Figure 6). Thus, flight missions will be possible using maximum speed, which will enable covering a larger area without having to worry about speed constraints. It will also help to be more efficient with the batteries of the UAVs, as flight times will be shorter.

RESULTS AND DISCUSSION
In our experiment, five DSMs for the same study area with different spatial resolution were obtained from the UAV at different flight altitudes (50m, 100, 200, 350m, 500m). Inside the same area, 187 points were measured by (using a) total station. A comparative procedure has been established by subtracting the elevation value obtained by Total station from the elevation value obtained on the five DSM that were extracted from the five different flight altitudes missions . The assessment of the DSMs will be shown first, according to the resolution, and second, according to the accuracy.
Firstly, the results shows a notable decrease of point cloud density with increasing flight altitudes (Figure 7 and Table 1). The points cloud density varies from 1.43 points/m2 at FA=500m (which is sufficient to flat areas) to 156 points/m2 at FA=50m, ; which is very valuable for high slopes. The dense clouds are fundamental to obtain DSMs and Orthomosaics. In this context, the results show that DSM pixel sizes are equal to 8cm, 16cm, 32cm, 58cm, 84cm and Ground Sample Distance (GSD) of the orthomosaic are equal to 2cm, 4cm, 8.2cm, 15m, 21cm for flight altitudes of 50m, 100m, 200m, 350m and 500m respectively ( Table 1). The variation of GSD directly affects the texture sharpness as shown in ( Figure A2), since we lose the clarity of the image. In fact, at FA=50m the GCPs are very clear, at FA=100 m and FA=200m the GCPs remains quite clear and at FA= 350 m and FA= 500m the GCPs are almost impossible to be distinguished. In addition, variation in DSM resolution, affect the representation of terrain roughness ( Figure  A3).
Secondly, linear regression is observed from the correlation scatterplots of all the elevation values that have been used. The correlation was established between UAV DSM at diverse altitude flights, using the total station measurements (Figure 8).
By analyzing the results more deeply to understand the impact of the flight altitude on the different error positions, the RMSE above |30 cm| has been selected for each flight altitude ( Figure  A5). Remarkable errors were found on a snow cornice (Figure 7 and Figure 9) and on the steep slope under it; which highlights the importance of the density of the point cloud. Going deeper, a sample section is created along the study area ( Figure 9). This figure shows a sample section that is selected in the middle of the study area. (Figure 10) focuses on the representation of cornice graphics, blending the results of the total station and the UAV at diverse flight altitudes.
High flight altitudes see a decrease of the DSM resolution, which affects the accuracy. As shown in (Figure 10 and Figure  A5) the majority of errors above |30 cm| are located on the steep changing slope which is the cornice location. In this location, the pixel size of the DSM is very critical to represent the real shape of the snow. The low density point cloud causes the rasterization, and the profile will be shown as a staircase (FA=500m) instead of a curved shape (FA=50m). Therefore we can understand the high minimum and maximum errors shown in (Table 1 and Figure A6) according the flight altitude changes.

CONCLUSION
Five different UAV surveys of the snow cover elevation were done with varying altitudes of 50 m, 100 m, 200 m, 350 m, and 500 m. The optimal DSM resolution increased from 8 cm at 50 m high to 84 cm at 500 m high; with the point density decreasing from 156 points/sqm to 1.43 points/sqm. As expected, the flight altitude of 50 m gave the sharpest terrain texture, offered the best GCPs visibility and the best representation of relief in particular for the steepest slopes. However, higher altitude flights provided a larger field of view. Based on this study we suggest that a flight altitude of 100 m is the best trade-off between accuracy and coverage, especially because it is compatible with the European regulation, which permits UAVs weighing less than 25kg. With a RMSE of 14 cm, this configuration can be used to characterize the snow height distribution with sufficient accuracy over complex terrains in Mount Lebanon. Figure A1: The five missions' plans at different altitudes  The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B2-2020, 2020 XXIV ISPRS Congress (2020 edition) Figure A4: Histogram plot of elevation differences (errors) between measured elevation by total station and the DSM extracted by UAV at diverse altitudes flights Figure A5: location of error above |30 cm| as per each flight altitude Figure A6: Accuracy of DSM UAV accuracy extracted from multiple flight altitudes comparing to Total Station measurements.