Integrating UAS photogrammetry and Digital Image Correlation for high-resolution monitoring of large landslides

: This paper shows the results of applying high-resolution Unmanned Aircraft System (UAS) photogrammetric surveying on a large landslide. A real case study, where permanently installing GCPs could be complex, where natural shaped and formed land pose severe limitations in deploying ground targets with optimal geometric configuration. We analysed performances in terms of survey accuracy obtained by performing photogrammetric surveys through the Zenmuse P1 DJI optical camera and Phantom 4 Pro 2. In combination with DJI Matrice 300 UAS, the P1 camera allows direct georeferencing through GNSS observations in RTK mode. Photogrammetric surveys, performed through different georeferencing methods, have been compared. Several targets have been permanently installed on the ground over the maximum vegetation height to guarantee long-lasting reference over the years in the area, which is characterised by a diffuse short vegetation cov-erage. Multitemporal UAS surveys have been then compared using Digital Image Correlation (DIC) algorithms, and deformation maps have been produced. Afterwards, DIC results were compared with observations made by the GNSS ground-based permanent receivers resulting in a standard deviation of 0,077 m. Through results analysis, good accordance between ground-based GNSS observations and DIC analysis on the photogrammetric surveys have been identified for the same time span. To conclude, this type of landslide presents a moderate deformation speed; in such a case, effective deformations monitoring could be achieved using pseudo-direct georeferencing, which permitted a 0.24 m accuracy on the whole tested area.


Introduction
Performing UAS photogrammetric surveying for landslide monitoring is an emerging methodology that well fit both the need to have periodic surveys and the capability to reach secluded areas [1]- [4]. One of the most deployed remote sensing techniques for landslide monitoring has undoubtedly been RADAR SAR Interferometry [5]- [9]. Its flexibility, together with long-range, accuracy and wide field of view, posed RADAR technology, both airborne [10], [11] and ground-based [12]- [14], as a reference in landslide monitoring. With such a technique, building a map of deformations representative of a wide area and maintaining a range resolution of 10-3 m, even from kilometric distances. However, hi-cost is limiting access to the RARAD technology. More often, good results have been obtained through the integrated use of different techniques [15]- [17]. Handling such a monitoring system could be difficult for minor stakeholders. In this fold, the photogrammetry technique has become increasingly used in this criticality. For technical reasons, performing UAS photogrammetric survey allows for discontinuous temporal monitoring only, even in the case of tethered aircraft, which could guarantee a longer-lasting flying [18], [19]. Nevertheless, considering many landslides have a slow or moderate velocity [20], a monitoring revisiting period of some hours or days could be considered an excellent way to monitor deformative patterns efficiently. Some steps must be performed following a specific workflow to properly carry out a photogrammetric survey and obtain a consistent 3D topographic reconstruction. The infield activities must be followed by specific processing. Post-processing is usually carried out in some hours/days, often using an office workstation, due to the considerable number of pictures to be elaborated for a photogrammetric reconstruction. For this reason, the time span to obtain a photogrammetric reconstruction could be considered as a daily recurrence. Depending on the landslide size, the landslide velocity, and the element at risk, a photogrammetric campaign could be considered an effective tool to create a topographic survey of the landslide body. A monitoring campaign could be performed by a diachronic comparison between two or more topographic surveys. By applying the Interferometric RADAR technique, a deformation map of the viewed area can be easily obtained. This seems far from a photogrammetric survey that produces a topographic model of the viewed area after photogrammetric reconstruction. Usually, indirect georeferencing is applied to the photogrammetric reconstruction process to obtain good accuracy that can be consistently applied to a deformative process like a landslide. Obtaining a centimetric accuracy on a photogrammetric survey could be considered a feasible objective if a good camera, a sound Ground Control Point (GCP) networks are arranged, and a proper flight altitude and camera settings have been chosen. Even air stability, which can interact with three canopy or long grass, can influence the quality of the final topographic model. However, once a georeferenced topographic model has been produced within a given accuracy, it can be compared with another georeferenced topographic model of the same area taken with a photogrammetric survey made at a different time. After creating DEMs from the 3d topographic models, a Dem of Difference (DoD) could be realised to investigate volumes changes and profile variations [21]- [23]. However, a pixel-based approach is to be considered to generate a deformation map of the studied area. Digital Image Correlation (DIC) technique [24], [25] can be applied. During the last decades, DIC has been applied for various engineering purposes.
The wide range of scales and the adequate accuracy and resolution that DIC can guarantee make it a very versatile technique. It is indeed used for architectural and structures deformations [26], tensile stress on construction materials [27], restoration [28], from microscale [29] to large scale applications [30]. This method allows specific algorithms to recognise light intensity patterns corresponding to the reflected light from the undeformed and deformed object configurations. In this way, deformations of small subsets of the image, thereby the deformations of small subsets of the actual object surface, can be measured [31], [32]. DIC is usually performed on multitemporal images of the same scene [33]. A Landslide, as long as a picture of the same area and with a similar point of view to each other, is used, could be considered a heterogeneous reflective body and investigated with the DIC technique [34].
Some researchers applied DIC on ortomosaics elaborated from photogrammetric surveys [35]- [37] to obtain a deformation map of the landslide area, and good results have been obtained.
In order to foster some advancements, some new features have been added to the research line in this work. In particular, a large area of investigation has been selected, direct and indirect georeferencing has been performed for the photogrammetric reconstruction process; the GCP and Check Point (CKP) networks has been linked to the local GNSS permanent monitoring system. A 45 Mpx camera has been used to guarantee hiresolution pictures, maintaining a small Ground Sampling Distance, which, considering the considerable size of the surveyed landslide, allowed to maintain a good flying altitude and reduce the time of the survey. The small GSD also allows for obtaining an effective orthophoto mosaic resolution.
Camera settings as aperture, shutter speed, magnification, and iso have been determined to optimise the UAS survey considering the landslide's size and velocity. The expected primary outcome is a deformation map of the landslide. For a specific deformative process, an effective monitoring system is expected to perceive displacements and resolutions that sufficiently characterise the process. Using a low time resolution monitoring system to monitor a fast landslide could not be effective, and adopting a low spatial resolution to perceive small particle movements.
The Calita landslide, the investigated area, hosts a well-studied landslide. Several studies have been performed, and different monitoring systems have been installed during the last 20 years. Thanks to these previous studies, the velocity range is accurately known; the photogrammetric surveys carried out in this research have been tailored to fit the velocity of the monitored process, i.e. surveys time recurrence, GSD, orthophoto mosaic resolution.
Furthermore, a GNSS network has been observing the landslide movements for the last five years; its outcomes have been used as a reference control for the performed survey.
The presented research activity has been designed to give convenient indications to implement effective landslide deformation monitoring using UAS photogrammetry. Survey and analysis were carried out on a real case study, where permanently installing GCPs could be economically tricky for local or regional administrations and where natural shaped and formed land poses severe limitations in deploying ground targets with optimal geometric configuration.

Study area
The Ca'Lita landslide, which is inside the administrative boundary of Baiso and into Reggio Emilia Province (Italy), was chosen as the study area for this research (Figure 1). The landslide is in the Secchia River Valley and is located in the Nord-Est mountainside of the northern Apennines [38]. Ca'Lita has a longitudinal length of approximately 3 km and a maximum width of about 1.4 km [39]. It is a large compound landslide composed of a roto-translational rockslide in the head zone, with flysch rock masses, and in the downslope has a translational earth slide-earthflow acting clayey complexes and debris material from the degradation of flysch rock masses [40].
The landslide of Ca'Lita has a long succession of reactivations that have threatened the safety of infrastructures and people in the crowning and valley area. In April 2004, an initial paroxysmal reactivation caused significant retrogression and advancement of the landslide [34].

Materials and Methods
The presented work executed photogrammetric surveys performed applying different georeferencing methods. In the study area, which is characterised by a diffuse short vegetation coverage, several targets have been permanently installed on the ground, over the maximum vegetation height, to guarantee a lasting reference on the ground over the years. Subsequently, photogrammetric flight missions were performed using nadiral and inclined camera angles of view. The installed Ground targets, used as GCPs and CkPs, were measured with GNSS observations in RTK and NRTK mode, both with a permanent local base station and a Virtual Local Station. Then the picture datasets were processed through photogrammetric software. The photogrammetric reconstruction processes were carried out using three different georeferencing methods. "A" method consists of a pseudo-direct georeferencing, in which coordinates of camera positions were directly provided by the onboard UAS GNSS receiver in RTK mode and then processed through SfM software. "B" method consists of an indirect georeferencing procedure, in which GCPs and CkPs coordinates have been observed through GNSS in field survey with a receiver in RTK mode. Through "C" method GCPs coordinates have been observed through GNSS in field survey with a receiver in RTK mode; furthermore, camera positions were directly provided by the onboard UAS GNSS receiver in RTK mode and then processed through SfM software.
The accuracy obtained through the A, B, C modes were investigated for three selected areas: the upper part of the landslide, so-called "Piana", the central part of the landslide, so-called "Colata", and the whole landslide area. Processing multitemporal UAS surveys were finally compared using Digital Image Correlation (DIC) algorithms and deformation maps have been produced. Afterwards, DIC results were compared with observations made by the GNSS ground based permanent receivers. A diachronic comparison of surveys produced in October 2020 and April 2021 was also performed and compared with observations made by the permanent GNSS monitoring system.

Photogrammetric surveys
Through photogrammetric surveys, orthomosaics have been generated using SfM Metashape software. Performing two photogrammetric campaigns and then producing orthomosaics at different times allows for comparing, identifying differences, and monitoring surface deformations. Monitoring Ca'Lita landslide outperformed two photogrammetric campaigns at two different times, considering these phenomena' high spatial and temporal variability.

UAS and Cameras
During photogrammetric surveys, two different drones and cameras were used. In October 2020, the DJI Phantom 4, a commercial-grade multi-copter drone with an integrated optical camera (DJI, 2016). DJI Matrice 300 RTK has been used for the April 2021 survey. It is an industrial aircraft equipped with RTK technology for a centimetre position (Dilectro, 2020). It integrates a 45MP full-frame sensor within the Zenmuse P1 camera, a global mechanical shutter, and an interchangeable fixed-focus DJI dl lens in a three-axis stabilised gimbal (DJI, 2020). Tables 1 and 2 below compare the equipment used. An automatic flight mode programmed through UgCS Software (Engineering, 2019) ( Figure 1) has been used in October and April survey campaigns. UgCS permits to set the specifications of drones and cameras in addition to Ground Sample Distance (GSD), side and forward overlap and camera tilt. Using planned flights allows to maintain speed, overlapping, and GSD reasonably constant along with the same photogrammetric campaign, and guarantees reliable repeatability for the various campaigns at different times, even months or years later. The characteristics of the flights are shown in Table 3.   GNSS surveys have been carried out, in October 2020 and April 2021, for targets georeferencing and to switch from a relative frame to an absolute reference system. 36 ground targets ( Figure 6) were installed on the study area, through deep-rooted poles ( Figure 6), as reference points for photogrammetric reconstruction. They were also used to materialise a long-lasting reference for multitemporal GNSS monitoring. Some permanent GNSS stations, working for more than 6 years in the study area, have been physically linked to some ground targets. This rigid junction allows for having an accurate longlasting ground truth. Each target was measured with the RTK mode and Network Real-Time Kinematic (NRTK), namely through Italpos service by Leica Geosystems (leica-geosystems.com). The receiver is an EMLID REACH RS2 multifrequency; the base receiver is an ItalPos service virtual base. In the second survey, in April 2021, the drone has an integrated RTK system, and a pseudo-direct survey has been carried out with the ItalPos virtual base service. The target position has been cquired with an EMLID REACH RS+ coupled with a permanent local base to obtain greater accuracy, Table 4 compares the specifications of the two EMLID receivers.   In this work, the software Agisoft Metashape 1.6.5 version has been used to process images acquired by UAS. The study area has been considered composed of a whole area ( ), the upper part, the so-called "Piana" ( ) and the central part, so-called "Colata" (Error! Reference source not found.). For each area, a photogrammetric project has been created. The surveys have been performed in April 2021, and in October 2020, each area has two projects, apart from the whole area surveyed during the April 2021 campaign only. For each project, the images have been imported without camera specifications and have been filtered following a quality threshold as the first step. The software then estimated interior and exterior parameters using EXIF georeferencing information. GCPs and CkPs were measured and manually selected on the project images as a second step. For the whole area, Colata and Piana, 36 targets were selected. In particular, 12 CkPs and 24 GCPs. The GCPs were then selected as a constraint during the bundle block adjustment (BBA) procedure to put the photogrammetric reconstruction within a local coordinate system. CkPs was only selected as Control points. In order to easily integrate the performed survey with other technical maps, the European Terrestrial Reference System ETRS89 in planar representation TM32 on the Terrestrial Reference Frame ETFR200 has been chosen. Once the bundle adjustment processes had been performed, exterior and interior camera parameters were adjusted accurately. A comparison between GCPs and CkPs model coordinates and the coordinates observed by the GNSS survey has been performed to assess georeferencing process accuracy. The accuracy has been expressed in pixels and meters. Root mean square error has been calculated for the GCPs and CkPs better to depict the error distribution in the overall study area. GCPs and CkPs have been installed in the Piana and the Colata only.

Digital Image Correlation (DIC)
DIC is an optical-numerical technique to quantify surface displacement. These deformations are calculated by comparing coregistered digital images collected at different epochs (Pan et al., 2008).
The correlation step in this process is essential and is based on searching for the maximum (or minimum) degree of correlation between the intensities within a reference region called the reference window (Pan et al., 2008). It is crucial to choose a suitable window's size; if it is too small, it will increase the noise of the image; instead, if it is too large, it may lose important details.
For this research, was used IRIS software owned by NHAZCA S.r.l. (Nhazca.it). The software wants as input a raster, and it was employed two orthomosaics with a resolution of 5 cm/px related to October 2020 and April 2021 derived from photogrammetric processing. In IRIS, performing pre-processing operations to standardise the images (called Master and Slave) and facilitate the detection of the movement is possible. After pre-processing, a Displacement Analysis with Phase Correlation (PC) algorithm was applied. The PC considers only the phase information but not the image content and base its functioning on the translation property of Fourier transform (FT). FT enunciate that "a shift of two relevant images in the spatial domain is transformed into the frequency domain as phase differences" [42]. The correlation function of Phase Correlation can be expressed through the inverse Fourier Transform of the normalised cross-power spectrum matrix Q: (1) In formula (1) F(u, v) and G(u, v) are FT of f(x, y) and g(x, y) (the functions of two images), F-1 is the inverse FT, i = √−1, Δx and Δy represent the shifts along with the two directions [41].
The PC method allows easy identification of the correlation peak given that the inverses FT of the normalised cross-power spectrum matrix Q is equivalent to Diarc delta function (formula 2) [41], [43].
Some spatial filters have been applied at the end of the process to decrease the noise and improve visualisation. In figure 6 it is shown the result of the DIC process. The displacement values of DIC processing were compared with the targets' position to evaluate the analysis results (Ground Control Point, GCP) in the two epochs considered, April 2020 and October 2021 (Error! Reference source not found.).

Results
Results from the photogrammetric project analysis are presented, particularly Colata, end the whole landslide area have been considered individually. Digital Image Correlation analyses and comparison with GNSS monitoring campaigns are reported for the Colata area.

Photogrammetric Reconstruction
Considering surface irregularities spread all over the study area, each flight mission has been performed using nadiral and crossed (+ 60°, -60°) views to optimise camera calibration and minimise shadowed areas. On the "whole area", the nadiral view only has been used to minimise the acquisition time. In table Error! Reference source not found., RMSE for each of the performed projects is reported. As far as method B for the whole area region is concerned, the vast extension of the study area, the consistent presence of vegetation put several issues to the SfM software. The photogrammetric reconstruction did not succeed.

Digital Image Correlation
The main product from Digital Image Correlation is a deformation map (Error! Reference source not found.). The analysis has been carried out on the "Colata" in the landslide medium part. The deformation map shows a central part in which surface deformation is visible. In particular, the flow direction is oriented from the northwest to South-East.  Table 6. Error! Reference source not found. shows GCPs positions observed in April 2021 (green) and October 2020 (pink). The mean square deviation is 0.077 m.

Discussion
In this study, the integration of photogrammetric surveys with GNSS permanent local monitoring allowed for investigating the effectiveness of photogrammetry as a tool for monitoring land deformations in complicated, vast and risk-prone areas. Direct and indirect georeferencing play a crucial role in determining the survey's accuracy. The whole studying area, of about 0.9 Km2, is also covered by vegetation, mostly grass from 10 to 60 cm tall (70%), some rushes area (2%) from 0,5 to 1,5 meters tall, and very sporadic areas with tall vegetation (< 0,3%).
In such a context, deploying a reliable and robust monitoring system could be a costly initiative, both in terms o time and money. Furthermore, maintaining a complex monitoring system as a RADAR SAR or multiple Total Station networks could be too demanding for small administrations like local municipalities. In this framework, the integration of detailed photogrammetric surveys, GNSS monitoring and Digital Image Correlation analyses have been exploited to investigate performances of a monitoring system based on ortophotomosaics from UAS photogrammetric campaigns.
Looking at the survey accuracy from photogrammetric reconstruction, comparing results from the April 2021 campaign carried out with Matrice 300 with P1 camera configuration, a lower RMSE is achieved with the indirect georeferencing procedure. The obtained value of 3 cm is almost one order of magnitude lower than that obtained through direct georeferencing, 24,4 cm.
By analysing results, an unexpected result needs to be interpreted. In particular, the C method should bring to best performances in terms of accuracy. On the contrary, as reported in Table 7, the photogrammetric reconstruction showed a better accuracy using method B. Theoretically, a more significant number of references should bring to a most accurate reconstruction. Georeferencing procedure for the C method is made with an infield survey only; on the contrary, the B method merges the in-field GNSS survey with the onboard GNSS RTK observations. However, in this case, despite the onboard GNSS having brought some more information to be digested by the SfM software, the final result does not comply with the expectation of finding better accuracy. The difference of about 1.5 cm could also be credited to the different base stations used for the GNSS in-field survey and onboard. For the B method, a GNSS permanent local station has been used. The permanent GNSS local station is a single frequency, and the onboard GNSS receiver is double frequency; using the permanent local base as RTK base for the onboard UAS receiver has not been possible. The cause of this unexpected result is most likely to be addressed to the hi-precision but low accuracy that virtual GNSS NRTK stations often guarantee. According to [44]) landslide can be classified in terms of velocity.  [34], [45]), the deformative process on the study area could be considered a slow to moderate velocity, classes 3-4. Applying the "A" method, which allows for 24 cm of accuracy, allows for effective monitoring performances on the whole area, guaranteeing a minimum time recurrency from 1 to 6 months; measuring a 50 cm displacement is plausible with an accuracy of 24 cm.
In the "colata" area, where 8 aligned targets were installed, the A, B and C methods also allow for adequate monitoring performances. Especially for the B case, with an accuracy of less than 4 cm, it allows for a minimum time recurrency up to some hours in case of a moderate deformation velocity.

Conclusions
The presented paper investigated how a photogrammetric campaign can perform deformation monitoring on a landslide body in a real case study. On a large landslide with compound deformative processes, 36 permanent ground targets have been installed through deep-rooted poles. Considering the considerable size of the landslide area, which is 0.9 Km2, deploying Ground Targets (GCPs and CkPs) with optimum geometrical distribution to achieve a proper photogrammetric survey is a very tough task. Furthermore, the area is strewn with marshy areas and a diffuse presence of hi-risk zones with collapsing material and unstable gullies. A considerable amount of investments should be provided to set up a proper photogrammetric survey with an optimised GCPs network in such a landslide area. This study shows that even with a few GCPs, integrating photogrammetric survey, Digital Image Correlation analysis, and GNSS monitoring can achieve effective deformation monitoring by applying indirect georeferencing. With a worse overall accuracy (0.24 m), pseudo-direct georeferencing may also produce sound indications on the deformation pattern. GNSS permanent stations give reliable, long-lasting knowledge on local deformation speed. Photogrammetry allows extending to the overall area GNSS observations with a valuable consistency. Finally, Digital Image Correlation analysis allows a detailed deformation map, a proper tool for interpreting deformations and landslide dynamics. Funding: This research received no external funding Data Availability Statement: Data is available on request from the authors. The data supporting this study's findings are available from the corresponding author, [Mugnai F.], upon reasonable request.