EVALUATION OF AN UNMANNED AERIAL VEHICLE (UAV) FOR MEASURING AND MONITORING NATURAL DISASTER RISK AREAS

This work presents an evaluation of a small UAV-RPAS-PPK to be used in the generation of digital surface models (DSM), without the need for control points, having as main application the monitoring of disaster risk areas (landslide and flooding). The areas to be measured are difficult to access, which prevents or makes access to the land difficult. In this evaluation, a study area of approximately 13 km2 was flown over, an average pixel of 11.6 cm and a total of 417 photos. The equipment used to acquire the images was a SenseFly eBee X, equipped with GNSS PPK for Direct Georeferencing (DG) and a camera model S.O.D.A. In all, 42 ground checkpoints were measured using a dual-frequency GNSS receiver. For both the measurement of the checkpoints and for the Direct Georeferencing (DG) base of the Unmanned Aerial Vehicles (UAV), a relative processing was performed, using the Brazilian Network for Continuous Monitoring (RBMC) as a reference. With this evaluation, it was possible to achieve a result (RMSE) for phototriangulation better than 1.2 pixels for horizontal and 1.5 pixels for vertical, without the need to measure any control points on the ground.


INTRODUCTION
Due to climate change, one of the biggest problems today is the occurrence of extreme weather events such as storms, floods, and landslides. These, in turn, combined with inadequate and disorderly urban occupation, cause great economic impacts and risks to people's lives. Hydrogeological impacts due to floods and landslides are among the most significant disasters in the world (Marta et al., 2020).
For the mapping and monitoring of disaster risk areas, there is a need for an updated database with its known qualities, in a context of rapid changes due to human activities and occupation and the climate events to which some areas are susceptible. It is also important to consider that in many risk areas, it is difficult or even impossible to access the study site for on-site mapping measurements, and there may be a risk to the people involved in this activity.
Unmanned Aerial Vehicles (UAV) data have been widely used for high resolution mapping and are presented as a low-cost alternative to photogrammetric acquisitions involving conventional platforms and cameras. In this sense, threedimensional geospatial data has the potential to provide essential information for decision-making (Fazeli et al., 2016) and to support disaster mitigation, prevention, and response strategies.

* Corresponding author
With the popularization of the UAV technology, several applications are found in the literature involving the issue of disasters related to landslides and floods (or hydrogeological disasters). As examples, analysis of road surface deformation caused by a landslide (Cardenal et al., 2019), identification of flooded areas (Giordan et al., 2017), and forensic geomorphological analysis to characterize a landslide-triggered debris flow (Cabral et al., 2021). However, there are few works that present the cartographic quality of the products generated from UAV data. Chiarando et al. (2019) state that the geospatial component in UAV data shall be properly considered through fast but rigorous photogrammetric processing aimed at generating 3D models and orthoimages of the surveyed areas. methodology for monitoring risk areas using Digital Surface Models (DSMs) produced using UAV technology.
In the development of the work, two surveying were carried out. The first was a risk area in the City of Blumenau, in the State of Santa Catarina, using a SenseFly eBee Classic UAV. According to the characteristics of this equipment, there was necessary to acquire control points on the ground, which made the work very difficult due to the impossibility of accessing various areas of the terrain (Reiss et al., 2018). The second surveying, which is the main discussion topic of this work, was carried out in the City of Santos, State of São Paulo, with a SenseFly eBee X UAV. Despite the statement of the manufacturer that there would be no need to acquire control points on the ground, up to a horizontal accuracy varying between 1 to 3 times the Ground Sampled Distance (GSD) (Roze et al, 2014), it is necessary to verify if in practice this statement and that quality of the products to be generated is true. Thus, where Ground Control Points (GCP) acquisition is impeded by site access difficulties or hazardous circumstances, Direct Georeferencing (DG) can be used (Ekaso et al., 2020).
Based on this statement, the hypothesis to be verified is whether it is possible to obtain a positioning quality as recommended by the manufacturer for without the use of ground control points. In this sense, this work aims to evaluate the quality of the 3D point cloud generated from data obtained with the SenseFly eBee X considering different photogrammetric processing configurations and the possibility of not needing the use of GCP.

Platform
The platform used in this work is the previously mentioned equipment, the SenseFly eBee X, with the ability to perform DG through an integrated dual-frequency GNSS, which allows the correction of the position of the images during or after the acquisition. The sensor used was a S.O.D.A. The main sensor characteristics are presented in Table 1 The equipment operating software is named eMotion 3, which allows both the planning and execution of photogrammetric coverage. Several functionalities are available, allowing different modes of equipment operation and coverage configurations of the area of interest. It is intuitive to use, favoring operation by lay users, with a minimum of training.

Case Study
The area selected for the case study is a place of real interest for CEMADEN. It is in the city of Santos, in the State of São Paulo, Brazil, as shown in Figure 1.
The main area of interest, represented in Figure 1 by the central polygon in blue, contains approximately 2.3 km². This area comprises the Morros de Santos, an area of high risk of landslides, especially in extreme rainfall events. The area of interest ( Figure 1) has regions that are difficult to access and some of them dominated by criminal factions, which makes it difficult or even impossible to measure control and check points.
As an alternative to this problem, a larger area, represented in Figure 1 by the pink polygon, was defined as an area of interest for acquiring the photos. The idea was to measure checkpoints around the area of interest, not within it, to avoid the risk area, so we ended up surveying a larger area. The GSD planned for the acquisition of photographs was 12 cm, but in view of the variation in relief in the study area, the average GSD performed was 11.6 cm. With this size of GSD the flight height was approximately 530 m. Being a risk area also because of crime, this height of flight made it possible for the UAV to perform a flight without being noticed by the local residents and/or by someone who could shoot it down. The lateral and longitudinal overlaps were defined as 60% and 75%, respectively. Table 2 summarizes the main parameters established for the flight plan.

Between photos 109 m
Flight line spacing 263 m Table 2. Summary of flight plan data.
The flight plan parameters were entered into the eBee X operating software, the eMotion 3 software and the photo acquisition were performed ( Figure 2). The result of the acquisition, which was carried out on August 23, 2019, generated a set of 417 photographs, on 5 flights in a total of 15 flight tracks.
A base station was measured at an arbitrary location, close to the landing and take-off point, for the Post-Processing Kinematic (PPK) positioning process of the acquired photos.
As this is an evaluation work, an effort was made, despite the difficulties mentioned, to measure checkpoints. Of the 43 points measured on the ground, 42 remained after the elimination of 1, as it was identified as containing a gross error. The survey of the checkpoints was carried out in 4 days, between the 13th to the 15th and on the 25th of November 2021. This time lapse of 2 years and three months between the acquisition of the photographs and the measurement of the checkpoints was due to logistical difficulties caused by the worldwide pandemic of covid-19.
To obtain the best possible quality for phototriangulation adjustment, it was considered to distribute the points, as much as possible, regularly spaced between them and covering the entire area of interest. Another important feature was that the points were photoidentifiable, as in corners for example.
Another arbitrary basis was established for the measurement of checkpoints. The measurement of the coordinates of the bases, both for the flight and for the checkpoints, was performed by tracking GNSS data, using dual-frequency Leica Viva GS-15 RTK GNSS receivers. Figure 2. Flight plan simulation run by eMotion: eBee X operating software.

GNSS Processing
The processing of the DG coordinates of the photos acquired with the UAV used is like what was performed for the measurement of checkpoints. In both cases, it is necessary to establish a reference base for the relative processing of GNSS coordinates ( Figure 3). The difference is that for the coordinates of the photos the positioning is the Kinematic Relative and in the case of the checkpoints it is the Static Relative (Cledat et al., 2020). In both cases, there was no need to materialize the base coordinates on the ground. They were calculated for the antenna phase center. The GNSS receiver was installed close to the survey region and the processing was carried out using coordinate transport through the stations of the Brazilian Network for Continuous Monitoring (RBMC) of the GNSS System (Figure 3), available in the national territory and whose data can be accessed by the website of the Brazilian Institute of Geography and Statistics (IBGE, 2022).
The RBMC reference stations that were used to calculate the base for the DG, named DG_BASE, are: POLI, EACH and UBA1 (IBGE, 2022). Table 3 shows the coordinates that were adjusted:  Table 3. Adjusted base of DG photos Once this base was calculated, the coordinates of the Perspective Centers (PC) of each of the 417 photos that were acquired in the photogrammetric survey with SenseFly eBee X were calculated using PPK, through the eMotion software in its PostFlight module. As it is a large volume of data, these data will not be shown here.

Photogrammetric Processing
The photogrammetric processing was performed in this work exclusively using the software Agisoft Metashape Professional Educational Edition, version 1.8.1. The first procedure performed was to insert the photos and assign to them the positions of the PCs calculated by the PPK processing performed in eMotion PostFlight and made available in a file. In the sequence, an automatic alignment (phototriangulation) of the photograph was performed, considering only the PCs as a control point. Bearing this in mind, a low-quality orthophoto was produced to allow locating photoidentifiable points, possible candidates for checkpoints to be located and measured on the ground. This occurred the day after the photo acquisition was carried out as soon as the processing of the base coordinates for the DG was ready. 43 points were then selected, of which 1 had gross errors and was eliminated. Left 42 points that were considered as tie points in a new phototriangulation process. This was performed, now considering better quality condition for the image processing parameters.

Experiment conditions
To define a better quality for the accomplishment of this work, some experimental combinations involving setup forms of photogrammetric processing was made: Setup 1: 1. Consider the coordinates of the GNSS base as a fixed (absolute injunction) in the DG's post-processing (Figure 4 (a)).
2. Consider the coordinates of the GNSS base as a relative variable (relative injunction) in the post-processing of the DG (Figure 4 (b)). Setup 2: 1. Perform a new alignment of the photogrammetric block (phototriangulation/optimize cameras) without automatic detection of new tie points after manually measuring the checkpoints in the photos ( Figure 5).
2. Perform a new alignment of the photogrammetric block (phototriangulation) but detecting new automatic tie points ( Figure 5).
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B2-2022 XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France Figure 5 -Selection of the accuracy (number of points) that will be automatically detected as tie points for the alignment process (image matching and phototriangulation).
Setup 3: 1. Use a photogrammetric block in which its accuracy was defined as high in its alignment ( Figure 5).
2. Use a photogrammetric block in which its accuracy was defined as low in its alignment ( Figure 5).
In 1 or 2, it was noticed that there was a difference in the quality of the phototriangulation result when it was used to fix the base coordinates for the DG or inform the base values and inform the Rinex file for the eMotion PostFlight module to recalculate the value of the base as a relative injunction.
Another change noticed in the results was after measuring the checkpoints in the photos (which work as tie points in the phototriangulation) that there was a change in the result if the automatic process of tie points detection by correspondence and a new phototriangulation were performed again.
Finally, it is known that the number of points detected for the image matching process interferes with the phototriangulation result. In this situation, it was chosen to test the accuracy of a High or Low point, seeking a better cost-effective processing ratio.

Statistical metrics
The verification of the quality of phototriangulation occurred by comparing the coordinates of the checkpoints measured in the photos (m) and calculated for the terrain by phototriangulation and the coordinates of their equivalent measured on the ground itself by means of GNSS (g). For this, the discrepancies between the points were then calculated (1): where: where: x d  is the mean of the discrepancies; x d  is the standard deviation of the sample; n is the sample size; x d RMSE is the root-mean-square error; x t is the t-Student statistic for sample; x 2  is the chi-square statistic for sample; e  is the standard deviation expected for the sample.

Considering
and being the theoretical t-Student (7) and chi-square (8) functions, respectively, for a significance level α, is possible calculate (9):

Results
To analyze whether it is possible to eliminate the need for control points, an analysis of the discrepancies between the values from the phototriangulation and those measured on the ground was performed. For this, it was necessary to define some input parameters for the statistical analysis. These parameters are shown in Table 6. From these input data, the analysis data presented in Table 7 were calculated. After the phototriangulation analysis was performed, an orthophoto of the photogrammetric block was produced, as shown in Figure 6.  Table 7 -Result for statistical analysis using different setups of photogrammetric processing Figure 6 -Printing of the orthophoto generated for the study area.

Analysis
In all, 8 experiments were carried out with processing parameters varying according to the description contained in Section 2.5. and the indication in the second column of Table 7. The objective was not to perform and present all possible combinations, as this would be impossible, but to present some variations that were shown to have a more significant influence on the results for the output coordinates of the 42 measured checkpoints. But regardless of any of the setups of the experiments we did, the largest RMSE was less than 1.9 in pixel, which was in experiment 5 in the East coordinate.
The best result achieved was experiment 8, having passed the trend test for all coordinate components, since it is true.
In the accuracy test, experiment 8 was also the best result, being ( ) also true.
In experiment 8 still time a horizontal RMSE less than 1.12 pixels and vertical better than 1.5 pixels.

CONCLUSIONS
The results presented in this case study allowed us to verify that the DG applied in data acquired using the SenseFly eBee X can generate products, as a dense cloud point that represents de DSM, with high quality without the need to measure any GCP.
Different configurations for triangulation processing were tested and the best result was achieved using configuration 8, which considers the following conditions: the coordinates of the GNSS base as a relative variable (relative injunction) in the postprocessing of the DG; a new alignment of the photogrammetric block (phototriangulation) but detecting new automatic tie points; and using a photogrammetric block in which its accuracy was defined as high in its alignment.
This result was achieved considering the kinematic relative positioning for the CP correction of the photos and the relative static positioning for the coordinates of the checkpoints. Some other settings for the photogrammetric process will be studied and applied to find ways to improve the results, such as exploring ways to improve the camera calibration consideration. This has not been aborded in this study.
Another subject that can be addressed in future works is the influence of connection points manually measured by a human operator on the final quality of the generated photogrammetric products.
We will now use the photogrammetric products that can be generated for mapping the study area to create a quality DSM and DEM that can be used in flood and landslide simulations.