Identifying Geomorphological Changes of Coastal Cliffs Through Point Cloud Registration from UAV Images

remote sensing

Article Identifying Geomorphological Changes of Coastal Cliffs through Point Cloud Registration from UAV Images

Xiangxiong Kong †

Department of Physics and Engineering Science, Coastal Carolina University, P.O. Box 261954, Conway, SC 29528-6054, USA; [email protected] or [email protected] † Formerly: Division of Civil Engineering, School of Engineering, University of Guam, UOG Station, Mangilao, GU 96913, USA.

Abstract: Cliff monitoring is essential to stakeholders for their decision-making in maintaining a healthy coastal environment. Recently, photogrammetry-based technology has shown great successes in cliff monitoring. However, many methods to date require georeferencing efforts by either measuring geographic coordinates of the ground control points (GCPs) or using global navigation satellite system (GNSS)-enabled unmanned aerial vehicles (UAVs), significantly increasing the implementation costs. In this study, we proposed an alternative cliff monitoring methodology that does not rely on any georeferencing efforts but can still yield reliable monitoring results. To this end, we treated 3D point clouds of the cliff from different periods as geometric datasets and further aligned them into the same coordinate system using a rigid registration protocol. We examined the performance of our approach through a few small-scale experiments on a rock sample as well as a full-scale field validation on a coastal cliff. The findings of this study would be particularly valuable for underserved coastal communities, where high-end GPS devices and GIS specialists may not be easily accessible resources. Citation: Kong, X. Identifying Geomorphological Changes of Keywords: cliff monitoring; UAV; photogrammetry; point cloud registration; iterative closest point; Coastal Cliffs through Point Cloud coastal erosion Registration from UAV Images. Remote Sens. 2021, 13, 3152. https://doi.org/10.3390/rs13163152 1. Introduction Academic Editors: Ana Nobre Silva Monitoring coastal cliffs is essential for maintaining a healthy coastal ecosystem and is and Cristina Ponte Lira particularly crucial for the island of Guam. Being the largest island in the Marianas Chain in the Western Pacific, Guam has a coastline of 125.5 km and 59% of it is rocky coastlines Received: 6 July 2021 characterized by steep cliffs and uplifted limestone terraces [1]. Due to the actions of Accepted: 5 August 2021 Published: 9 August 2021 the sea, strong winds, ground motions, and water surges [2], coastal cliffs are prone to erosion. For example, Typhoon Halong in 2002 struck Guam and led to erosion on the

Publisher’s Note: MDPI stays neutral southeast shorelines; the 1993 Guam earthquake (magnitude of 7.8) also caused slides in with regard to jurisdictional claims in coastal cliffs throughout the island [3]. Other natural impacts such as seasonal changes on published maps and institutional affil- rock thermal stress and/or cliff vegetation could also influence the cliff stability and cause iations. geological hazards. Cliff erosions could lead to sediments on coastal reefs and weaken the integrity of a local coastal ecosystem. One engineering approach to address this concern is to monitor the cliff erosion process using advanced technologies, based on which results can be delivered to the stakeholders for making timely decisions in managing a coastal zone. Traditionally, Copyright: © 2021 by the author. Licensee MDPI, Basel, Switzerland. cartographic geological mapping [4,5] is the most popular method for surveying coastal This article is an open access article erosion. However, this method is labor-intensive and prone to error due to mapping distributed under the terms and inaccuracy [6]. In addition, field deployments at inaccessible locations could be challenging conditions of the Creative Commons and time-consuming. As such, terrestrial laser scanning (TLS)-based technology [7,8] has Attribution (CC BY) license (https:// received increasing attention in coastal surveying for being able to achieve a non-contact creativecommons.org/licenses/by/ and accurate solution through creating dense 3D point clouds of coastal areas. Nevertheless, 4.0/). the laser scanner could be costly, and inconvenient for field deployment due to its heavy

Remote Sens. 2021, 13, 3152. https://doi.org/10.3390/rs13163152 https://www.mdpi.com/journal/remotesensing Remote Sens. 2021, 13, 3152 2 of 17

self-weight. In addition, the scanner must be operated by trained technicians. Hence, field cliff monitoring on a routine basis may not be easily achievable. Recently, photogrammetry-based methods have shown great successes in coastal monitoring. Utilizing computer vision algorithms, 2D images of the coastal cliffs can be processed to create dense 3D point clouds. Through the usages of unmanned aerial vehicles (UAVs), images of inaccessible coastal areas can be obtained. For example, Westoby et al. [6] created a 3D model of a coastal cliff in northeast England using a photogrammetry-based workflow. The authors concluded that the discrepancy between the photogrammetry model and the one obtained by TLS technology was less than 0.04 m. Letortu et al. [7] performed a similar investigation by comparing the datasets from TLS, terrestrial photogrammetry, and UAV photogrammetry using a testbed in France. Hayakawa and Obanawa [9] applied UAVs and photogrammetry to detect the volumetric changes of a coastal cliff on the Suzume-Jima Island in Japan. The authors of [10–12] review the recent work of photogrammetry-based remote sensing methods in coastal mapping. A commonality from the above work is that georeferencing is required. This allows point clouds collected at different periods to be aligned into the same geographic coordinate system so that the geomorphological changes of the cliff can be identified. To date, the most dominant georeferencing approach is based on ground control points (GCPs). Made by large-size artificial targets, ideally, GCPs shall be evenly deployed to the cliff area. Georeferencing coordinates of each GCP then are measured by global positioning system (GPS) devices (e.g., total stations [7], global navigation satellite systems (GNSSs) [13,14]). GCPs-based georeferencing has several limitations. First, many locations of the cliff area could be inaccessible for deploying and measuring GCPs. For instance, a steep cliff face would be dangerous for GCP deployments due to rock falls [7]. Second, GPS measurements of GCPs usually require an error that is less than a few mm or cm, requiring high-cost GPS measuring devices (e.g., more than $10,000). Lastly, geographic information system (GIS) specialists need to be hired for the field GPS data collections and processing which would bring the extra cost to the project. Some researchers have proposed direct georeferencing methods without using GCPs. The idea is to directly mount customized GNSS modules to portable digital cameras [15,16] or onboard UAV cameras [17,18]. As a result, the geotagged photos collected in the field can assist in the 3D mapping of a coastal cliff with accurate geographic coordinates. Methods based on direct georeferencing can map inaccessible areas of the cliff where deploying GCPs could be dangerous. However, methods based on direct georeferencing are still costly and would require extensive investigations on GNSS hookups [17]. Most recently, efforts have been made by investigating professional UAVs equipped with onboard off-the-shelf GNSS receivers (e.g., DJI Phantom 4 real-time kinematic (RTK) [19,20]). Nevertheless, the accuracy of elevation generated from the DJI Phantom 4 RTK could be problematic [21]. In this manuscript, we present a non-georeferenced approach for coast cliff monitoring. We apply a photogrammetry workflow to reconstructed dense 3D point clouds of the cliff from different periods. Then, different point clouds are aligned into the same coordinate system through a rigid registration protocol such that the geomorphological changes of the cliff can be identified. We examine the performance of our approach through a series of small-scale experiments on a rock sample against different lighting conditions and surface textures. It is followed by a full-scale field validation on a coastal cliff in Guam. Thereafter, we discuss the results from both small-scale and full-scale validations. The major contribution of this study is to propose an alternative cliff monitoring methodology that does not rely on any georeferencing efforts but can still yield reliable monitoring results. Many previous researchers have focused on cliff monitoring using GCPs and/or GNSS-enabled UAVs. These georeferencing-based methods would be inflexi- ble due to the high cost of hiring GIS specialists and/or purchasing expensive hardware. In contrast, our approach solely relies on computational algorithms to align point clouds together for uncovering the geomorphological changes caused by cliff erosion, significantly reducing the implementation cost. Although point cloud processing techniques such as Remote Sens. 2021, 13, x FOR PEER REVIEW 3 of 17

Remote Sens. 2021, 13, 3152 3 of 17 together for uncovering the geomorphological changes caused by cliff erosion, significantly reducing the implementation cost. Although point cloud processing techniques thesuch iterative as the iterative closest pointclosest (ICP) point algorithm (ICP) algorithm [22,23] [ have22,23] beenhave previouslybeen previously investigated investi- in cliffgated monitoring in cliff monitoring [24–28], [2 the4–28 roles], the ofroles the of ICP the ICP algorithm algorithm in these in these studies studies are are limited limited as supplementalas supplemental tools tools in improvingin improving point point cloud cloud alignment alignment accuracy accuracy within within the the georeferencing georefer- framework.encing framework. To the bestTo the knowledge best knowledge of the of author, the author, there isthere no literatureis no literature developing developing a com- pletelya completely non-georeferenced non-georeferenced cliff monitoring cliff monitoring methodology. methodology. The findings The findings of this of studythis study would bewould particularly be particularly valuable valuable for Guam for Guam and other and underservedother underserved coastal coastal communities, communities, where high-endwhere high GPS-end devices GPS devices and trained and trained GIS professionals GIS professionals may not may be easilynot be accessible easily accessible resources. resources. The rest of this manuscript is organized as follows: Section2 illustrates the research The rest of this manuscript is organized as follows: Section 2 illustrates the research methodology and explains the technical details; Section3 demonstrates the soundness methodology and explains the technical details; Section 3 demonstrates the soundness of of the proposed method through a series of small-scale experiments; Section4 validates the proposed method through a series of small-scale experiments; Section 4 validates the the method using a full-scale coastal cliff; Section5 further discusses applicability and method using a full-scale coastal cliff; Section 5 further discusses applicability and limita- limitations of our method; and Section6 concludes the study. tions of our method; and Section 6 concludes the study. 2. Methodology 2. Methodology The research methodology, illustrated in Figure1, contains three major components The research methodology, illustrated in Figure 1, contains three major components that include (a) image collection, (b) point cloud reconstruction, and (c) point cloud reg- that include (a) image collection, (b) point cloud reconstruction, and (c) point cloud regis- istration. Our method starts with the image collection of the cliff using UAVs. Then, tration. Our method starts with the image collection of the cliff using UAVs. Then, UAV UAV images are further processed by a series of computer vision algorithms, termed images are further processed by a series of computer vision algorithms, termed structure- structure-from-motion with multi-view stereo (SfM-MVS), to reconstruct the point cloud from-motion with multi-view stereo (SfM-MVS), to reconstruct the point cloud of the cliff. of the cliff. Next, a new point cloud of the cliff can be obtained using the same procedure Next, a new point cloud of the cliff can be obtained using the same procedure after the after the second field visit. Thereafter, these two point clouds are aligned into the same second field visit. Thereafter, these two point clouds are aligned into the same coordinate coordinate system through a protocol of rigid registration, which contains a few computa- system through a protocol of rigid registration, which contains a few computational algo- tional algorithms for point cloud alignment. Finally, the differential changes between two rithms for point cloud alignment. Finally, the differential changes between two well- well-alignedaligned point point clouds clouds can be can extracted be extracted through through computing computing the cloud thecloud-to-cloudto-cloud distance. distance. As Asa result, a result, the the geomorphological geomorphological changes changes of ofthe the cliff cliff can can be beidentified. identified. Each Each component component in in thethe researchresearch methodologymethodology is further further explained explained in in the the rest rest of of this this section. section.

FigureFigure 1.1.Research methodology of of this this study study where where (a ()a image) image collection; collection; (b) ( bpoint) point cloud cloud reconstruc- reconstruc- tion; and (c) point cloud registration. tion; and (c) point cloud registration.

2.1. Image Collection A large volume of digital images of the target cliff are collected using UAVs (see Figure1a). Many consumer-grade UAVs can fit such a role. The flight routes and camera parameters (e.g., ISO, shutter speed, image resolution, and camera shooting interval) can be predefined through built-in flight operation apps. UAV images are intended to cover Remote Sens. 2021, 13, 3152 4 of 17

the cliff with different camera positions and angles. Adjacent images shall have enough overlapping for matching feature points that will be explained in Section 2.2.

2.2. Point Cloud Reconstruction UAV images are processed by SfM-MVS for creating a 3D point cloud of the cliff (see Figure1b). SfM-MVS is a well-established photogrammetry workﬂow that has been widely applied to coastal surveying [29], civil infrastructure inspection [30], river bathymetry extraction [31], and historic building preservation [32]. To this end, feature points (i.e., tie points, key points), which are small image patches that contain unique intensity distributions, are detected from each UAV image. Because feature points are invariant against image translation, rotation, and scaling, feature points with similar intensity distributions can be consistently tracked and matched across multiple UAV images. Some of the well- known features are scale-invariant feature transform (SIFT) [33], Shi-Tomasi [34], features from accelerated segment test (FAST) [35], Harris-Stephens [36], binary robust invariant scalable keypoints (BRISK) [37], and speeded up robust features (SURF) [38]. Next, feature points across different UAV images are matched based on their levels of similarities in intensity distributions. A geometric transformation matrix is also estimated in this stage to describe the relations between matched feature pairs (i.e., correspondences) of two adjacent UAV images. Based on the transformation matrix, incorrect matching results (i.e., outliers) can be eliminated. Thereafter, SfM algorithms are adopted to estimate both extrinsic parameters (e.g., locations and orientations) and intrinsic parameters (e.g., focal length and pixel sensor size) of the camera. The 3D geometry of the cliff scene is also calculated in this stage. Then, camera positions and angles are further reﬁned through bundle-adjustment algorithms to reduce reprojection errors in MVS. Next, multiple-view UAV images and their corresponding camera parameters are utilized for reconstructing the sparse 3D point cloud of the cliff. Users can also examine the quality of reconstruction errors in the sparse point cloud, and if needed, may change the parameters of the algorithms to re-create the sparse point cloud. Finally, pixels are back-projected to all UAV images to create an RGB-colored dense point cloud, which represents the 3D surface of the cliff. The detailed reviews of SfM-MVS are summarized in [39–41].

2.3. Point Cloud Registration To uncover the geomorphological changes of the cliff, two dense point clouds at different periods are aligned together using the protocol of rigid registration (Figure1c ). The protocol can find geometric similarities of two point clouds and applies rotation, scaling, and translation to rigidly align one point cloud to another. This procedure further contains three steps that include (1) scaling one point cloud to a real-world length unit; (2) rough alignment of two point clouds based on manually selected correspondences; and (3) fine alignment of two point clouds using the automated ICP algorithm. Each step is further explained as follows. As shown in Figure1c, point cloud A is first scaled to the correct real-world unit using a scaling factor, which is the ratio of the distance between two existing points measured from the cliff site in the real world over the distance of the same two points from the point cloud. The point cloud after scaling is considered as the reference point cloud which will not move for the rest of the registration procedure. Then, point cloud B (denoted as the floating point cloud) is roughly aligned to the reference point cloud (i.e., point cloud A) through manually finding correspondences. Correspondences are points that appear at similar locations in both reference and floating point clouds. Selections of correspondences are flexible as long as they can be visually identified. Based on correspondences, a geometric transformation matrix can be estimated, allowing the floating point cloud to be rigidly translated, rotated, and scaled for matching the reference point cloud. Remote Sens. 2021, 13, x FOR PEER REVIEW 5 of 17 Remote Sens. 2021, 13, x FOR PEER REVIEW 5 of 17 Remote Sens. 2021,, 13,, xx FORFOR PEERPEER REVIEWREVIEW 5 of 17

Correspondences are points that appear at similar locations in both reference and floating Remote Sens. 2021, 13, 3152 Correspondences are points that appear at similar locations in both reference and floating5 of 17 pointCorrespondences clouds. Selections are points of correspondences that appear at similar are flexible locations as inlong both as reference they can andbe visuallyfloating pointCorrespondences clouds. Selections are points of correspondences that appear at similar are flexible locations as inlong both as reference they can andbe visually floating identified.point clouds. Based Selections on correspondences, of correspondences a geometric are flexible transformation as long as they matri canx canbe visually be esti- identified.point clouds. Based Selections on correspondences, of correspondences a geometric are flexible transformation as long as they matri canx can be visually be estimated,identified. allowing Based the on floating correspondences, point cloud a to geometric be rigidly transformation translated, rotated, matri andx can scaled be esti- for mated,identified. allowing Based the on floating correspondences, point cloud a to geometric be rigidly transformation translated, rotated, matri andx can scaled be esti- for matchingmated,Due allowing to the the reference manual the floating point selection cloud.point ofcloud correspondences, to be rigidly translated, errors are rotated, inevitably and scaled introduced for matchingmated, allowing the reference the floating point cloud.point cloud to be rigidly translated, rotated, and scaled for duringmatchingDue rough tothe the alignment.reference manual point selection Such cloud. errors of correspondences, can be further reduced errors are through inevitably fine introduced registration. matchingDue tothe the reference manual point selection cloud. of correspondences, errors are inevitably introduced Hereduring weDue rough adopt to the alignment. the manual ICP algorithm selectionSuch errors to of further correspondences,can be optimize further reduced the errors transformation arethrough inevita finebly matrix. registration. introduced The ICP duringDue rough to the alignment. manual selectionSuch errors of correspondences,can be further reduced errors through are inevita finebly registration. introduced Hereduring we rough adopt alignment. the ICP algorithm Such errors to further can be optimize further thereduced transformation through fine matrix. registration. The ICP algorithmHereduring we rough adopt starts alignment.the with ICP an algorithm initial Such guess errors to further of can the be rigidoptimize further body thereduced transform transformation through of two fine matrix. point registration. clouds,The ICP and algorithmHere we adopt starts the with ICP an algorithm initial guess to furtherof the rigid optimize body thetransform transformation of two point matrix. clouds, The andICP iterativelyalgorithmHere we adopt improvesstarts thewith ICP the an transformationalgorithm initial guess to furtherof matrixthe rigid optimize through body thetransform repeatedly transformation of findingtwo point matrix. correspondences clouds, The andICP iterativelyalgorithm startsimproves with thean initialtransformation guess of the matrix rigid throughbody transform repeatedly of two finding point correspond- clouds, and withiterativelyalgorithm minimum startsimproves errors. with the an The initialtransformation last rowguess of of Figure the matrix rigid1c illustratesthrough body transform repeatedly comparisons of two finding ofpoint two correspond- clouds, point cloudsand encesiteratively with minimumimproves theerrors. transformation The last row matrixof Figure through 1c illustrates repeatedly comparisons finding ofcorrespond- two point atencesiteratively each with stage minimumimproves of the registration. errors.the transformation The last row ofmatrix Figure through 1c illustrates repeatedly comparisons finding ofcorrespond- two point cloudsences with at each minimum stage of errors. the registration. The last row of Figure 1c illustrates comparisons of two point cloudsencesThe with at rough each minimum stage alignment of errors. the registration. can The effectively last row of align Figure two 1c illustrates point clouds comparisons together of but two small point mis- cloudsThe at rough each stage alignm of entthe canregistration. effectively align two point clouds together but small misa- alignmentscloudsThe at rough each may stage alignm exist. of entthe Fine canregistration. alignment, effectively onalign the two other point hand, clouds is capabletogether ofbut adjusting small misa- small lignmentsThe rough may exist.alignm Fineent alignment, can effectively on the align other two hand, point is clouds capable together of adjusting but small small misa- mis- misalignmentslignmentsThe rough may but exist. alignm may Fineent not alignment, can work effectively well on if thethe align initialother two hand, misalignment point is cloudscapable oftogether of two adjusting point but cloudssmall small misa- mis- is large. alignmentslignments may but mayexist. not Fine work alignment, well if theon theinitial other misalignment hand, is capable of two of point adjusting clouds small is large. mis- Byalignmentslignments successively may but mayexist. adopting not Fine work alignment, these well two if the on alignments initialthe other misalignment hand, in the is correct capable of two order, of point adjusting the clouds misalignments small is large. mis- Byalignments successively but mayadopting not work these well two if thealignments initial misalignment in the correct of order two point, the misalignmentsclouds is large. betweenByalignments success twoively but point mayadopting clouds not work these can well be two gradually if alignmentsthe initial reduced. misalignment in the correct of order two point, the misalignmentsclouds is large. betweenBy success twoively point adopting clouds canthese be two gradually alignments reduced. in the correct order, the misalignments betweenBy success twoively point adopting clouds canthese be two gradually alignments reduced. in the correct order, the misalignments 3.between Small-Scale two point Validation clouds can be gradually reduced. 3. Small-Scale Validation 3.1.3. Small Test Configuration-Scale Validation 3.1.3. Small Test -ConfigurationScale Validation 3.1. ATest series Configuration of small-scale tests on a rock sample was performed with the purposes of 3.1. TestA series Configuration of small- scale tests on a rock sample was performed with the purposes of (1) (1) reconstructingA series of small dense-scale 3D tests point on a clouds rock sample from thewas testperformed sample with under the differentpurposes lightingof (1) reconstructingA series of dense small -3Dscale point tests clouds on a rock from sample the test was sample performed under with different the purposes lighting of and (1) andreconstructing surfaceA series texture of dense small conditions; -3Dscale point tests clouds andon a (2)rock from detecting, sample the test was localizing, sample performed under and with different quantifying the purposes lighting differential of and (1) surfacereconstructing texture denseconditions; 3D point and clouds(2) detecting, from the localizing test sample, and underquantifying different differential lighting fea-and featuressurfacereconstructing texture of the rockdenseconditions; sample 3D point and under clouds(2) detecting, geometric from the localizing changes.test sample, and To underquantifying this end,different a differential rock lighting sample fea-and was turessurface of texturethe rock conditions; sample under and (geometric2) detecting, changes. localizing To ,this and end, quantifying a rock s ampledifferential was col-fea- collectedturessurface of texturethe from rock Tumon conditions; sample Bay under inand Guam (geometric2) detecting, in June changes. 2020.localizing TheTo ,this longestand end, quantifying diametera rock s ampledifferential of the was sample col-fea- is lectedtures offrom the Tumon rock sample Bay in under Guam geometricin June 2020. changes. The longest To this diameter end, a ofrock the s sampleample wasis about col- aboutlectedtures 13.5 offrom the cm, Tumon rock as shownsample Bay inin underGuam figure geometricin (a) June in 2020. Table changes. The1. Five longest To test this diameter cases end, were a ofrock establishedthe s sampleample wasis to about mimiccol- 13.5lected cm from, as shown Tumon in Bay figure in Guam (a) in inTable June 1. 2020. Five Thetest longestcases were diameter established of the tosample mimic is differ- about different13.5lected cm from, testingas shown Tumon environments. in Bay figure in Guam (a) in The Tablein thirdJune 1. 2020. columnFive Thetestof caseslongest Table were1 diameter elaborates established of the the to sample different mimic is differ- about lighting ent13.5 testing cm, as shownenvironments. in figure The (a) inthird Table column 1. Five of test Table cases 1 wereelaborate establisheds the different to mimic lighting differ- conditionsent13.5 testing cm, as and environments.shown geometric in figure The changes (a) inthird Table forcolumn 1. each Five of test test Table case.cases 1 Thewereelaborate rock establisheds sample the different to in mimic Case lighting differ- A had a conditionsent testing andenvironments. geometric changesThe third for column each test of Tablecase. The1 elaborate rock samples the differentin Case Alighting had a darkerconditionsent testing texture andenvironments. due geometric to the high changesThe moisturethird for column each content test of Tablecase. after The1 the elaborate rock sample samples the was differentin collected Case Alighting fromhad a the darkerconditions texture and due geometric to the highchanges moisture for each content test aftercase. theThe sample rock sample was collected in Case fromA had the a beach.darkerconditions Images texture and of due geometric Cases to the B tohigh changes E were moisture takenfor each content a few test days aftercase. later; theThe sample hence,rock sample was the samplecollected in Case has fromA a had brighter the a beach.darker Imagestexture of due Case to sthe B to high E were moisture taken acontent few days after later the; hence sample, the was sample collected has a from brighter the surfacebeach.darker Images texture.texture of due Case tos the B to high E were moisture taken acontent few days after later the; hence sample, the was sample collected has a frombrighter the surfacebeach. Images texture. of Cases B to E were taken a few days later; hence, the sample has a brighter surfacesurface texture.texture. surfaceTable texture. 1. Test matrix for the small-scale validation. Table 1. Test matrix for the small-scale validation. Table 1. Table 1. Test matrix for the small-scale validation. Table 1. Test matrix for the small-scale validation. Sample and Test TestTest Case Surface Surface Texture Lighting Condition Geometric Change Sample and Test Test Surface Lighting Condition Geometric Change SampleEnvironment and Test CaseTest TextureSurface Lighting Condition Geometric Change SampleEnvironment and Test Case Texture Lighting Condition Geometric Change Environment Case Texture Environment Indoor lighting condition; Indoor lighting condition; daylight was the CaseCase A ADark Indoor Dark lighting condition; daylightdaylight was was the the only ReferenceReference dataset dataset Case A Dark onlyIndoor light lighting source condition; daylight was the Reference dataset Case A Dark onlyIndoor light lighting source condition; daylightlight was source the Reference dataset Case A Dark only light source Reference dataset only light source

Outdoor lighting condition;Outdoor the sample lighting was condition; Case B Light Outdoor lighting condition; the sample was No geometricNo geometric change was change made CaseCase B BLight directlyOutdoor Light under lighting sunlight condition; the the sample sample was was directly No geometric change was made Case B Light directlyOutdoor under lighting sunlight condition; the sample was No geometric change was made Case B Light directly under sunlight No geometric changewas made was made directly under sunlight under sunlight

Indoor lighting condition; daylight was the Case C Light Indoor lighting condition;Indoor daylight lighting was the condition; Three small stones were added (see Figure 2a) Case C Light onlyIndoor light lighting source condition; daylight was the Three small stonesThree were small added stones (see were Figure 2a) CaseCase C CLight onlyIndoor light Light lighting source condition; daylightdaylight was was the the Three only small stones were added (see Figure 2a) Case C Light only light source Three small stonesadded were (see added Figure (see2 a)Figure 2a) only light source light source

Indoor lighting condition; the roof lamp A thin layer of salt particles was added Case D Light Indoor lighting condition; the roof lamp A thin layer of salt particles was added Case D Light wasIndoor the lighting only light condition; sourceIndoor the roof lighting lamp condition; A thethin layerA thin(seeof salt layerFigure particles of 2b) salt was particles added Case D Light wasIndoor the lightingonly light condition; source the roof lamp A thin layer(see of salt Figure particles 2b) was added CaseCaseRemote D D Sens.Light 2021 , 13was, x FOR the Light onlyPEER lightREVIEW source roof lamp was the only (see Figurewas added 2b) 6 of 17 was the only light source (see Figure 2b) light source (see Figure2b)

Outdoor lighting condition; A thin layer of salt particles Outdoor lighting condition; the sample was A thin layer of salt particles was added to a CaseCase E E Light Light the sample was placed in was added to a different placed in the shadow different location (see Figure 2c) the shadow location (see Figure2c)

Figure 2. Bird’s-eye views of the rock sample to show the geometric changes under different test cases. (a) three stones were added in Case C; (b) a thin layer of salt particles was added in Case D; and (c) a thin layer of salt particles was added in Case E at a different location.

A consumer-grade digital camera (Sony Alpha 6400 with the E PZ 16–50 mm Lens) was adopted for image collection. The auto mode was selected to allow the camera to define its preferred shooting parameters. The distance between the lens and the rock sample varied from 20 to 40 cm during image collection. Images were shot with a resolution of 6000 pixels by 4000 pixels. In Cases A to E, 199, 86, 70, 67, and 98 images were collected, respectively.

3.2. Point Cloud Reconstruction The 3D point clouds of the sample were reconstructed using the off-the-shelf software Agisoft Metashape (version 1.6.2) [42] installed on a mobile workstation (Lenovo ThinkPad P72 with 16 GB of RAM and a 2.2 GHz CPU). Here, we use Case A as an example to illustrate the workflow. Figure 3a shows 7 out of 199 images of the rock sample in Case A. The collected images were then aligned together through the SfM-MVS algorithm. A sparse point cloud (Figure 3c) is first constructed, based on which a dense point cloud is built, as shown in Figure 3d. The camera positions are estimated in Figure 3b where small blue patches indicate the camera positions and angles.

Figure 3. A 3D reconstruction of the rock sample in Case A. (a) Sample input images; (b) camera positions; (c) sparse point cloud; and (d) dense point cloud.

Remote Sens. 2021, 13, x FOR PEER REVIEW 6 of 17

Outdoor lighting condition; the sample was A thin layer of salt particles was added to a Case E Light placed in the shadow different location (see Figure 2c)

Remote Sens. 2021, 13, x FOR PEER REVIEW 6 of 17

To mimic the landscape changes that one would see in a cliff, some geometric features of the rock sample were intentionally changed in Cases C, D, and E (see the fourth column of Table 1). Briefly, in Case C, three small stones denoted S1, S2, and S3 were placed on Outdoor lightingthe top condition; of the rock the sample was (see A Figure thin layer 2a). of Insalt Case particles D, insteadwas added of to adding a stones, a thin layer Case E Light of salt particles was added on the top of the sample (see Figure 2b). Thereafter, such a Remote Sens. 2021, 13, 3152 placed in the shadow different location (see Figure 2c) 6 of 17 layer was removed, and a new layer of salt particles was added to a different location of the sample in Case E (see Figure 2c). To mimic the landscape changes that one would see in a cliff, some geometric features of the rock sample were intentionally changed in Cases C, D, and E (see the fourth column of Table 1). Briefly, in Case C, three small stones denoted S1, S2, and S3 were placed on the top of the rock sample (see Figure 2a). In Case D, instead of adding stones, a thin layer of salt particles was added on the top of the sample (see Figure 2b). Thereafter, such a layer was removed, and a new layer of salt particles was added to a different location of the sample in Case E (see Figure 2c).

FigureFigure 2. Bird 2. ’Bird’s-eyes-eye views views of the of the rock rock sample sample to to show show the the geometricchanges changes under under different different test test cases. cases. (a) three (a) three stones stones were wereadded added in Case in Case C; (b C;) (ab thin) a thin layer layer of ofsalt salt particle particless was was added added inin CaseCase D; D; and and (c )(c a) thina thin layer layer of salt of salt particles particle wass addedwas added in Casein CaseE at a E different at a different location. location.

A consumerTo mimic the-grade landscape digital changes camera that (Sony one would Alpha see 6400 in a cliff,with some the E geometric PZ 16–5 features0 mm Lens) wasof adopted the rock samplefor image were collection. intentionally The changed auto inmode Cases was C, D, selected and E (see to the allow fourth the column camera to Figure 2. Bird’s-eye viewsdefine ofof the Table its rock preferred1). sample Briefly, to inshooting show Case the C, geometricparameters. three small changes stones The under distance denoted different S1, between S2, test and cases. the S3 were(lensa) three placedand stones the on rock the sam- were added in Case C; (bple) a topthin varied oflayer the from of rock salt 20 sampleparticle to 40s (seewas cm Figureaddedduring in2a). Caseimage In D; Case andcollection. D, (c) instead a thin layerImages of adding of salt w particleere stones, shots was a with thin added layera resolution of in Case E at a different location.salt particles was added on the top of the sample (see Figure2b). Thereafter, such a layer of 6000 pixels by 4000 pixels. In Cases A to E, 199, 86, 70, 67, and 98 images were collected, was removed, and a new layer of salt particles was added to a different location of the respectively.sampleA consumer in Case E- (seegrade Figure digital2c). camera (Sony Alpha 6400 with the E PZ 16–50 mm Lens) was adoptedA consumer-grade for image digitalcollection. camera The (Sony auto Alpha mode 6400 was with selected the E PZto allow 16–50 the mm camera Lens) was to 3.2.defineadopted Point its Cloud for preferred image Reconstructio collection.shootingn parameters. The auto mode The wasdistance selected between to allow the thelens camera and the to rock define sam- its plepreferredThe varied 3D shooting frompoint 20 clouds parameters.to 40 cm of duringthe The sample image distance werecollection. between reconstructed Images the lens w andere using theshot rock thewith sampleoff a -resolutionthe varied-shelf soft- wareoffrom 6000Ag 20isoft pixels to 40 Metashape cmby during4000 pixels. image (version In collection. Case 1.6.2)s A to Images[ E,42 199,] installed were86, 70, shot 67, on withand a mobile 98 a resolution images workstation were of 6000 collected, pixels (Lenovo by 4000 pixels. In Cases A to E, 199, 86, 70, 67, and 98 images were collected, respectively. ThinkPadrespectively. P72 with 16 GB of RAM and a 2.2 GHz CPU). Here, we use Case A as an exam- ple3.2. to illustratePoint Point Cloud Cloud the Reconstructio Reconstruction workflow.n Figure 3a shows 7 out of 199 images of the rock sample in Case A. The collected images were then aligned together through the SfM-MVS algorithm. The 3D pointpoint cloudsclouds of of the the sample sample were were reconstructed reconstructed using using the the off-the-shelf off-the-shelf software soft- A sparse point cloud (Figure 3c) is first constructed, based on which a dense point cloud wareAgisoft Ag Metashapeisoft Metashape (version (version 1.6.2) [ 421.6.2)] installed [42] installed on a mobile on a workstation mobile workstation (Lenovo ThinkPad (Lenovo is builtThinkPadP72 with, as shown 16P72 GB with ofin 16 RAMFigure GB andof 3RAMd. a 2.2The and GHz ca a mera2.2 CPU). GHz positions Here,CPU). we Here are use, estimatedwe Case use ACase as in anA Figure as example an exam- 3b towhere smallpleillustrate blueto illustrate patches the workflow. the indicate workflow. Figure the 3Figure cameraa shows 3a positions 7shows out of 7 199 out and images of angles.199 ofimages the rock of the sample rock insample Case A.in CaseThe collected A. The collected images wereimages then were aligned then togetheraligned together through thethrough SfM-MVS the SfM algorithm.-MVS algorithm. A sparse Apoint sparse cloud point (Figure cloud3 c)(Figure is first 3 constructed,c) is first constructed, based on based which on a densewhich pointa dense cloud point is built,cloud isas built shown, as in shown Figure in3d. Figure The camera 3d. The positions camera arepositions estimated are inestimated Figure3 bin where Figure small 3b where blue patches indicate the camera positions and angles. small blue patches indicate the camera positions and angles.

Figure 3. A 3D reconstruction of the rock sample in Case A. (a) Sample input images; (b) camera positions; (c) sparse point cloud; and (d) dense point cloud.

Figure 3. A 3D 3D reconstruction reconstruction of the rock sample in Case A A.. ( a)) Sample input images; images; ( b) camera position positions;s; ( c) sparse point cloud; and (d) dense point cloud. cloud; and (d) dense point cloud.

Figure4 shows the 3D reconstruction results of Cases B to E. The dense point clouds of the sample have different surface colors due to changes in lighting conditions. For instance, the dense point clouds have a lighter color representation in Cases B and E (Figure4a,d) compared with dense point clouds in Cases C and D (Figure4b,c). This is because the

sample was in an outdoor environment for the former test cases. Additionally, notice that the dense point cloud in Case A (Figure3d) has a slightly darker color than Cases C and D Remote Sens. 2021, 13, x FOR PEER REVIEW 7 of 17

Remote Sens. 2021, 13, x FOR PEER REVIEW 7 of 17

Figure 4 shows the 3D reconstruction results of Cases B to E. The dense point clouds of theFigure sample 4 shows have differentthe 3D reconstruction surface colors resultsdue to changesof Cases in B lightingto E. The conditions. dense point For clouds in- ofstance, the sample the dense have point different clouds surface have a colorslighter duecolor to representation changes in lighting in Case sconditions. B and E (Figure For instance,4a,d) compared the dense withpoint dense clouds point have clouds a lighter in Case colors Crepresentation and D (Figure in 4 bCase,c). Thiss B and is because E (Figure Remote Sens. 2021, 13, 3152 4athe,d) sample compared was within an denseoutdoor point environment clouds in for Case thes formerC and testD (Figure cases. Additionally4b,c). This7 ofis, 17 noticebecause thethat sample the dense was point in an cloudoutdoor in Case environment A (Figure for 3d) the has former a slightly test darker cases. colorAdditionally than Case, notices C thatand the D (denseFigure point 4b,c). cloudThis is in caused Case Aby ( theFigure fact 3thatd) has Case a slightlyA has a higherdarker moisture color than content, Cases C anddespite (FigureD (Figure all4 b,c).three 4 Thisb ,testc). is Thiscases caused is being caused by the under fact by that theindoor Casefact lighting Athat has Case a higher conditions. A has moisture a higher content, moisture despite content, despiteall three all three test cases test beingcases under being indoor under lighting indoor conditions. lighting conditions.

Figure 4. A 3D reconstruction of the rock sample in (a) Case B; (b) Case C; (c) Case D; and (d) Case E.

Figure Figure4. A 3D 4. Areconstruction 3D3.3. reconstruction Point Cloud of the of theRegistration rock rock sample sample in in (a)) CaseCase B; B; (b ()b Case) Case C; (C;c) Case(c) Case D; and D; ( dand) Case (d) E. Case E. To align dense point clouds together, we adopt open-source software, CloudCom- 3.3.pare Point3.3. (version Point Cloud Cloud 2.10.2) Registration Registration [43], and first scale the point cloud in Case A with the real-world unit. To doTo th alignTois, align two dense densepoints point point (#4332244 cloudsclouds together,and together, #3697936 we adoptwe in adopt Figure open-source open 5a) were- software,source selected software CloudCompare in the, CloudCom- unscaled (version 2.10.2) [43], and ﬁrst scale the point cloud in Case A with the real-world unit. To do parepoint (version cloud. 2.10.2)The distance [43], and between first scale these the two point points cloud was in measured Case A with as 7.753 the real from-world Cloud- unit. this, two points (#4332244 and #3697936 in Figure5a) were selected in the unscaled point Compare. Notice that there is no real-world dimension associated with this distance. Next, To docloud. this, The two distance points between (#4332244 these and two #3697936 points was in measured Figure 5 asa) 7.753 were from selected CloudCompare. in the unscaled pointtheNotice locations cloud. that The of there these distance is notwo real-world pointsbetween were dimension these identified two associated points in the withwasrock thismeasuredsample distance. and as the Next, 7.753 corresponding the from loca- Cloud- Compare.distancetions of wasNotice these measured two that points there as were 10.5is no identiﬁed cm. real This-world in further the dimension rock led sample to aassociated and scaling the corresponding factor with of th 10.5is distance. distance cm/7.753 Next, = the1.354 locationswas cm/ measured1. Thereafter,of these as 10.5 two cm. the points This initial further were point led identified cloud to a scaling was in factorscaledthe rock of up10.5 sample by cm/7.753 multiplying and = the 1.354 corresponding 1.354 cm/1 to. the distancecoordinatesThereafter, was ofmeasured the each initial point. point as 10.5The cloud newcm. was Thispoint scaled further cloud up by, ledafter multiplying to scaling a scaling, 1.354 is treated factor to the of coordinatesas 10.5the referencecm/7.753 = pointof eachcloud. point. Figure The 5 newb illustrates point cloud, the aftercomparison scaling, isof treatedthe point as theclouds reference before point and cloud. after scal- 1.354 cm/1. Thereafter, the initial point cloud was scaled up by multiplying 1.354 to the ing.Figure 5b illustrates the comparison of the point clouds before and after scaling. coordinates of each point. The new point cloud, after scaling, is treated as the reference point cloud. Figure 5b illustrates the comparison of the point clouds before and after scaling.

FigureFigure 5. (a) 5.The(a) Theunscaled unscaled point point cloud cloud fro fromm CaseCase A; A; (b ()b a) comparisona comparison of unscaled of unscaled and scaled and pointscaled clouds point in clouds Case A; in and Case (c) the A; and (c) theunscaled unscaled point point cloud cloud in Casein Case C can C can be aligned be aligned to the to reference the reference point cloudpoint bycloud manually by manually selecting selecting four correspondences four correspondences (A0-R0,(A0-R0, A1-R1, A1-R1, A2-R2, A2-R2 and, and A3-R3). A3-R3). Figure 5. (a) The unscaled point cloud from Case A; (b) a comparison of unscaled and scaled point clouds in Case A; and Next, a point cloud from a new test case is aligned to the reference point cloud. We (c) the unscaled point cloud in CaseNext, C can a pointbe aligned clou dto fromthe reference a new test point case cloud is aligned by manually to the selecting reference four point correspond- cloud. We use Case C as an example here for illustration. First, rough registration was performed ences (A0-R0, A1-R1, A2-R2, useand CaseA3-R3). C as an example here for illustration. First, rough registration was performed usingusing four four correspondences correspondences (A0 (A0-R0,-R0, A1-R1,A1-R1, A2-R2, A2-R2 and, and A3-R3 A3-R3 in Figurein Figure5c) from 5c) from both both point clouds. Thereafter, ﬁne registration was conducted through the ICP algorithm. Point point clouds. Thereafter, fine registration was conducted through the ICP algorithm. Point cloudsNext, froma point Case clou B, D,d andfrom E werea new aligned test case with theis aligned point cloud to the in Case reference A using point the same cloud. We useclouds procedure,Case from C as Case but an the exampleB, proceduresD, and hereE were of thesefor aligned illustration. alignments with arethe First, notpoint shownrough cloud in registration thisin Case manuscript A using was due theperformed to same usingthe four length correspondences constraint. (A0-R0, A1-R1, A2-R2, and A3-R3 in Figure 5c) from both point clouds. Thereafter, fine registration was conducted through the ICP algorithm. Point clouds from Case B, D, and E were aligned with the point cloud in Case A using the same

Remote Sens. 2021, 13, x FOR PEER REVIEW 8 of 17

Remote Sens. 2021, 13, 3152 8 of 17

procedure, but the procedures of these alignments are not shown in this manuscript due to the length constraint. 3.4. Point Cloud Comparison 3.4. PointOnce Cloud point Comparison clouds of Case B to E are aligned with the reference point cloud in Case A, theOnce differential point clouds features of canCase be B identifiedto E are aligned through with computing the reference cloud-to-cloud point cloud distance in Case inA, CloudCompare. the differential features The cloud-to-cloud can be identified distance through between computing Case A–B cloud and-to- Casecloud A–C distance are illustratedin CloudCompare in Figure. 6The. As cloud shown-to- incloud the figure,distance the between test sample Case in A Case–B and B experiencedCase A–C are no il- geometriclustrated in change Figure but 6. As was shown under in a differentthe figure, lighting the test condition. sample in AsCase a result,B experienced the cloud- no to-cloudgeometric distance change betweenbut was Caseunder B a anddifferent reference lighting point condition. cloud (i.e., As Casea result, A) isthe extremely cloud-to- smallcloud (0.07 distanc cme inbetween Figure 6Casea,b), B indicating and reference two point cloud clouds (i.e. match, Case well A) is with extremely each other. small The(0.07 three cm in stones Figure in 6 Casea,b), indicating C can be identifiedtwo point fromclouds the match cloud-to-cloud well with each distance other. as The shown three instonesFigure in6 d,eCase. The C can locations be identified of stones from agree the well cloud with-to- thecloud ground distance truth as measurements shown in Figure in Figure6d,e. The6f. Furthermore,locations of stones the height agree ofwell S1, with S2, and the S3ground can be truth roughly measurements quantified in as Figure 0.4, 0.3, 6f. andFurthermore, 0.7 cm. the height of S1, S2, and S3 can be roughly quantified as 0.4, 0.3, and 0.7 cm.

FigureFigure 6. 6.( a(a,b,b)) Cloud-to-cloud Cloud-to-cloud distance distance between between Case Case A andA and B; (Bc); point(c) point cloud cloud in Case in Case B serves B serves as the as the ground truth measurement; (d,e,g,h,j,k) cloud-to-cloud distance between Case A and C; and ground truth measurement; (d,e,g,h,j,k) cloud-to-cloud distance between Case A and C; and (f,i,l) point (f,i,l) point clouds in Case C serves as the ground truth measurements. Results in (a,d,g,j) are in log clouds in Case C serves as the ground truth measurements. Results in (a,d,g,j) are in log scale with scale with the unit of cm; and results in (b,e,h,k) are in linear scale with the unit of cm. the unit of cm; and results in (b,e,h,k) are in linear scale with the unit of cm. The cloud-to-cloud distances between Case A–D and Case A–E are shown in Figure The cloud-to-cloud distances between Case A–D and Case A–E are shown in Figure7 . 7. As can be seen in the first and second columns of the figure, salt particles in the test As can be seen in the first and second columns of the figure, salt particles in the test samples samples in Cases D and E can be identified. The cloud-to-cloud distance in log scale has in Cases D and E can be identified. The cloud-to-cloud distance in log scale has better demonstrationsbetter demonstrations on finding on finding the boundary the boundary of the particles;of the particles; while thewhile result the inresult linear in scalelinear isscale more is suitablemore suitable for quantifying for quantifying the thickness the thickness of the of salt the layer. salt layer. Results Results indicate indicate that thethat proposedthe proposed method method can can reliably reliably find find geometric geometric changes changes that that occurred occurred in in the the test test sample, sample, regardlessregardless of of changes changes in in the the lighting lighting conditions, conditions, as as seen seen in in Cases Cases D D and and E. E.

RemoteRemote Sens. Sens. 2021 2021,, 13, 13,, 3152x, xFOR FOR PEER PEER REVIEW REVIEW 9 9of ofof 17 17

FigureFigure 7. 7. 7. ((a a(,b,ab,,bd,d,de,e),)e Cloud-to-cloud )Cloud Cloud-to-to-cloud-cloud distance distance distance between between between Case Case CaseA and A A D; and and (c, fD )D point; ; ( c (,cf,)f clouds)point point in cloud cloud Cases s Din in serves Case Case asD D servestheserves ground as as the the truth ground ground measurements; truth truth measurement measurement (g,h,j,k) cloud-to-clouds;s (; g(g,h,h,j,jk,k) )cloud cloud distance-to-to-cloud-cloud between distance distance Case between Abetween and E; Case andCase (Ai ,Al )and and point E E; ; andcloudsand ( i(,il,) inl )point point Case cloud Ecloud servess sin in asCase Case the E groundE serves serves as truth as the the measurements.ground ground truth truth measurement measurement Results in (as,d.s ,Results.g Results,j) are in in in ( log a(,ad,d, scaleg,g,j,)j )are withare in in log scale with the unit of cm; and results in (b,e,h,k) are in linear scale with the unit of cm. thelog unit scale of with cm; andthe unit results of cm; in ( band,e,h results,k) are inin linear(b,e,h, scalek) are with in linear the unit scale of with cm. the unit of cm.

4.4. Field Field Field Validation Validation Validation 4.1.4.1. Site Site Site Description Description Description AA cliff cliff at at Tagachang Tagachang Beach Beach in in Guam Guam is is selected selected as as the the testbed testbed for forfor field ﬁeldfield validation. validation. TagachangTagachang Beach Beach is is located located on on the the east east side side of of the the island. island. Showing Showing in in Figure Figure 88 d,8d,d, thethe the cliffcl cliffiff startsstarts at at at the the the south southsouth end endend of ofof the thethe beach beachbeach and and extends extends to to the the south. south. south. A A Asmall small small portion portion portion of of ofthe the the cliff cliff cliff is is selectedisselected selected in in inthis this this study study study (see (see (see the the the white white white circle circle in in Figure FigureFigure 8 8d).d). Figure Figure 99 illustrates9 illustrates illustrates the the the testbed testbed testbed from different views. The target cliff is about 30 m high measured from the cliff bottom fromfrom different different views. views. The The target target cliff cliff is is about about 30 30 m m high high measur measureded from from the the cliff cliff bottom bottom and has a relatively ﬂat top surface covered by vegetation (Figure9a). Both north and east andand has has a a relatively relatively flat flat top top surface surface covered covered by by vegetation vegetation ( Figure(Figure 9 9a).a). Both Both north north and and east east sides of the target cliff are steep rock surfaces (Figure9b,c). A rock slide can be observed sidessides of of the the target target cliff cliff are are steep steep rock rock surfaces surfaces ( Figure(Figure 9b 9b,c).,c). A A rock rock slide slide can can be be observed observed on the east vertical plane of the cliff due to the previous erosion (Figure9d). onon the the east east vertical vertical plane plane of of the the cliff cliff due due to to the the previous previous erosion erosion ( Figure(Figure 9 9d).d).

FigureFigure 8. 8. The The location location of of ofthe the the testbed. testbed. testbed. ( a()a )shows ( ashows) shows the the island the island island of of Guam. Guam. of Guam. ( b(b–d–d) ) (are bare– dblow )blow are-upblow-upup orthographic orthographic orthographic satellite satellite satellite views views accord- accord- views ingly. Tagachang Beach and the target cliff are illustrated in Figure 8d. All satellite images are generated from Google accordingly.ingly. Tagachang Tagachang Beach Beach and andthe thetarget target cliff cliff are are illustrated illustrated inin Figure Figure 88d.d. All All satellite images areare generatedgenerated from from Google Google Earth. Map data in (a,b) are from Google, NOAA, Maxar Technologies, and CNES/Airbus; map data in (c,d) are from Earth.Earth. MapMap datadata in in ( a(,ab,b)) are are from from Google, Google, NOAA, NOAA, Maxar Maxar Technologies, Technologies, and and CNES/Airbus; CNES/Airbus; mapmap datadata inin ( c(,cd,d)) are are from from GoogleGoogle and and CNES/Airbus. CNES/Airbus. Google and CNES/Airbus.

Remote Sens.Remote 2021 Sens., 13,2021 x FOR, 13, PEER 3152 REVIEW 10 of 17 10 of 17

Figure 9.Figure (a) Bird’s 9. (a)- Bird’s-eyeeye view view of the of thetarget target cliff; cliff; ( (bb) viewview from from north; north; (c) view(c) view from from the east; the and east; (d) blow-upand (d) viewblow of-up the view rock of the rock slide.slide. CP CP in inFigure Figure 99a is thethe center center point point for for UAV UAV ﬂight flight mode mode of POI. of POI.

4.2. UAV4.2. UAV Operation Operation,, Data Data Collection Collection,, andand Point Point Cloud Cloud Reconstruction Reconstruction TwoTwo visits visits were were carried carried outout on on 25 25 June June and and 11 July 11 2020,July 2020, respectively. respectively. The east sideThe ofeast side the cliff was inaccessible due to high tides during both visits. Hence, the deployment work of thewas cliff performed was inaccessible at the north sidedue ofto the high cliff tides (i.e., see during the deployment both visits. area Hence in Figure, the8d). deployment Two workoff-the-shelf was performed UAVs, at the the DJI north Air (SZ side DJI of Technology the cliff (i.e Co.,., Ltd,see Shenzhen,the deployment China) andarea DJI in Figure 8d). PhantomTwo off- 4the Pro-shelf + V2.0 UAVs, (DJI Phantom the DJI 4,Air hereafter, (SZ DJI SZ Technology DJI Technology Co., Co., Ltd, Ltd, Shenzhen, Shenzhen, China) and China),DJI Phantom were adopted 4 Pro as + tools V2.0 for (DJI image Phantom collection. 4, hereafter, SZ DJI Technology Co., Ltd, Shenzhen,To China evenly) capture, were theadopted testbed as under tools different for image camera collection. positions, two image collection strategies were proposed. The first strategy was to take a series of images under a pre- To evenly capture the testbed under different camera positions, two image collection programmed flight route to scan the cliff from the top. This was achieved by operating strategiesthe DJI were Air through proposed. an off-the-shelf The first strategy smartphone was app, to take Pix4Dcapture a series of (version images 4.10.0) under [44 a], prepro- grammedinstalled flight on an route iPhone to 11.scan A double-gridthe cliff from mapping the top. mission This was was created achieved in the by app. operating The the DJI Airaltitude through of the an flight off was-the defined-shelf smartphone as 90.2 m with app front, Pix4Dcapture and side overlapping (version of 90%4.10.0) and [44], in- stalled75%, on respectively, an iPhone based 11. A on double which- thegrid app mapping calculated mission the UAV was locations created for in shooting the app. each The alti- tudeimage. of the Asflight a result, was 83defined images as were 90.2 collected m with by front the DJIand Air side for overlapping both field visits of with90% anand 75%, image resolution of 4056 pixels by 3040 pixels. The UAV camera angle was selected as respectively, based on which the app calculated the UAV locations for shooting each im- 80 degrees. age. As aFor result, the second 83 images image were collection collected strategy, by the images DJI wereAir for captured both field by the visits DJI Phantom with an image resolution4 through of 4056 an intelligent pixels by mode, 3040 named pixels. point The ofUAV interest camera (POI), angle using was the selected smartphone as 80 app degrees. DJIFor Go the 4 second (version image 4.3.36) collect [45]. Theion appstrateg wasy, preinstalled images were in thecaptured all-in-one by DJIthe remoteDJI Phantom 4 throughcontroller. an Theintelligent POI mode mode, allowed named the UAV point to fly of alonginterest a circular (POI), path using horizontally the smartphone with a app DJI Gopredefined 4 (version center 4.3.3 point6) and[45]. a radius.The app The was center preinstalled point was defined in the at all the-in cliff’s-one top DJI (see remote the con- white cross in Figure9a), and the radius was selected as 62 m. Then, multiple POI flights troller.were The performed POI mode under allowed altitudes the of 25UAV m to to 45 m.fly Imagesalong werea circular automatically path horizontally collected by with a predefinedthe onboard center UAV point camera and using a radius. a camera The shooting center point interval was of 2defined seconds at with the an cliff’s image top (see the whiteresolution cross of 4864in Figure pixels by9a), 3648 and pixels. the Inradius total, 284was and selected 251 images as 62 were m. collectedThen, multiple in the POI flightsfield were visits performed of 25 June and under 11 July, altitudes respectively. of 25 m to 45 m. Images were automatically collected byFigure the onboard10a,c show UAV the samplecamera UAV using images a camera from shooting the DJI Phantom interval 4 underof 2 seconds the POI with an imagemode resolution for both fieldof 4864 visits. pixels Figure by10 b,d3648 show pixels. thecamera In total positions, 284 and where 251 theimages backgrounds were collected are sparse point clouds of the testbed. As can be seen in the figures, the DJI Air follows in theflight field missions visits of of 25 a 3-by-3June and grid 11 to July, cover respectively. the top of the cliff area. The DJI Phantom 4 isFigure operated 10a in,c POIshow mode the tosample mainly UAV scan theimages east and from north the sides DJI Phantom of the cliff 4 from under four the POI modedifferent for both altitudes. field visits. Figure 10b,d show the camera positions where the backgrounds are sparseBased point on theclouds collected of the UAV testbed. images As from can the be DJI seen Air and in the Phantom figures, 4, the the dense DJI pointAir follows flightclouds missions of two of field a 3 visits-by- are3 grid reconstructed to cover usingthe top Agisoft of the Metashape cliff area. on a workstationThe DJI Phantom (Dell 4 is XPS 8930-7814BLK-PUS with 32 GB of RAM and a 3.0 GHz CPU). Figure 11 illustrates operated in POI mode to mainly scan the east and north sides of the cliff from four differ- the dense point clouds from both field visits where point clouds outside the scope of the ent altitudes.testbed are truncated. The point cloud in the 25 June visit contains 48.5 million points, while the point cloud on 11 July contains 55.8 million points.

Remote Sens. 2021, 13, x FOR PEER REVIEW 10 of 17

Figure 9. (a) Bird’s-eye view of the target cliff; (b) view from north; (c) view from the east; and (d) blow-up view of the rock slide. CP in Figure 9a is the center point for UAV flight mode of POI.

4.2. UAV Operation, Data Collection, and Point Cloud Reconstruction Two visits were carried out on 25 June and 11 July 2020, respectively. The east side of the cliff was inaccessible due to high tides during both visits. Hence, the deployment work was performed at the north side of the cliff (i.e., see the deployment area in Figure 8d). Two off-the-shelf UAVs, the DJI Air (SZ DJI Technology Co., Ltd, Shenzhen, China) and DJI Phantom 4 Pro + V2.0 (DJI Phantom 4, hereafter, SZ DJI Technology Co., Ltd, Shenzhen, China), were adopted as tools for image collection. To evenly capture the testbed under different camera positions, two image collection strategies were proposed. The first strategy was to take a series of images under a prepro- grammed flight route to scan the cliff from the top. This was achieved by operating the DJI Air through an off-the-shelf smartphone app, Pix4Dcapture (version 4.10.0) [44], installed on an iPhone 11. A double-grid mapping mission was created in the app. The alti- tude of the flight was defined as 90.2 m with front and side overlapping of 90% and 75%, respectively, based on which the app calculated the UAV locations for shooting each image. As a result, 83 images were collected by the DJI Air for both field visits with an image resolution of 4056 pixels by 3040 pixels. The UAV camera angle was selected as 80 degrees. For the second image collection strategy, images were captured by the DJI Phantom 4 through an intelligent mode, named point of interest (POI), using the smartphone app DJI Go 4 (version 4.3.36) [45]. The app was preinstalled in the all-in-one DJI remote con- troller. The POI mode allowed the UAV to fly along a circular path horizontally with a predefined center point and a radius. The center point was defined at the cliff’s top (see the white cross in Figure 9a), and the radius was selected as 62 m. Then, multiple POI flights were performed under altitudes of 25 m to 45 m. Images were automatically collected by the onboard UAV camera using a camera shooting interval of 2 seconds with an image resolution of 4864 pixels by 3648 pixels. In total, 284 and 251 images were collected in the field visits of 25 June and 11 July, respectively. Figure 10a,c show the sample UAV images from the DJI Phantom 4 under the POI mode for both field visits. Figure 10b,d show the camera positions where the backgrounds are sparse point clouds of the testbed. As can be seen in the figures, the DJI Air follows Remote Sens. 2021, 13, 3152 flight missions of a 3-by-3 grid to cover the top of the cliff area. The DJI Phantom11 of4 17is operated in POI mode to mainly scan the east and north sides of the cliff from four different altitudes.

Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of 17

Figure 10. (a) Sample DJI Phantom 4 images from the visit on 25 June 2020; (b) camera positions from the visit on 25 June 2020; (c) sample DJI Phantom 4 images from the visit on 11 July 2020; and (d) camera positions from the visit on 11 July 2020.

Based on the collected UAV images from the DJI Air and Phantom 4, the dense point clouds of two field visits are reconstructed using Agisoft Metashape on a workstation (Dell XPS 8930-7814BLK-PUS with 32 GB of RAM and a 3.0 GHz CPU). Figure 11 illustrates the dense point clouds from both field visits where point clouds outside the scope Figure 10. (a) Sample DJI Phantom 4 images from the visit on 25 June 2020; (b) camera positions from the visit on 25 June of the testbed are truncated. The point cloud in the 25 June visit contains 48.5 million 2020; (c) sample DJI Phantom 4 images from the visit on 11 July 2020; and (d) camera positions from the visit on 11 July 2020. points, while the point cloud on 11 July contains 55.8 million points.

Figure 11. Dense point clouds of the cliff from (a) the 25 June visit; and (b) the 11 July visit. Correspondences are selected Figure 11. Dense point clouds of the cliff from (a) the 25 June visit; and (b) the 11 July visit. Correspondences are selected from both point clouds for rough alignment. from both point clouds for rough alignment. 4.3. Point Cloud Registration 4.3. Point Cloud Registration To align the point clouds, we first scale the point cloud into a correct real-world unit To align the point clouds, we first scale the point cloud into a correct real-world unit in CloudCompare. To this end, we treat the point cloud in the second visit on 11 July as in CloudCompare. To this end, we treat the point cloud in the second visit on 11 July as the reference point cloud. During this visit, three markers (M1, M2, and M3) were placed the reference point cloud. During this visit, three markers (M1, M2, and M3) were placed in the testbed as seen in Figure 12d. M1 and M2 were X marks made by the blue paint in the testbed as seen in Figure 12d. M1 and M2 were X marks made by the blue paint tape, while M3 was the UAV landing pad. The distances between the three markers were tape, while M3 was the UAV landing pad. The distances between the three markers were taken by a measuring tape (see the second column in Table 2). Next, the markers were taken by a measuring tape (see the second column in Table2). Next, the markers were visually identified from the dense point cloud (Figure 12a–c). The distances between three visually identified from the dense point cloud (Figure 12a–c). The distances between three markers in the point cloud were also measured (see the third column in Table 2). Finally, markers in the point cloud were also measured (see the third column in Table2). Finally, three scaling factors were calculated, based on which the average scaling factor of 1.054 m three scaling factors were calculated, based on which the average scaling factor of 1.054 m was applied for scaling the point cloud in the second visit. was applied for scaling the point cloud in the second visit. Thereafter, the scaled point cloud in the first visit was aligned to the reference point cloud through the registration protocol. Figure11 demonstrates the selections of correspondences (A1-R1, A2-R2, A3-R3, and A4-R4) from both point clouds for rough alignment. Next, the point cloud in the first visit was further aligned by the automated ICP algorithm. Figure 13 shows the comparison of point clouds under different views of the cliff from two visits during the registration procedure. The point cloud from the 25 June visit is rendered in blue. As can be seen in Figure 13b,e,h, small misalignments can be observed after rough alignment. Such misalignments can be minimized after fine alignment is performed (Figure 13c,f,i). Figure 12. Deployment of markers during the visit on 11 July where (a) shows the dense point cloud in the top view; (b) shows the blow-up detail of the point cloud with three markers; (c) shows blow-up details of three markers truncated from the dense point cloud; and (d) shows the images of markers collected by a smartphone camera in the field.

Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of 17

Figure 11. Dense point clouds of the cliff from (a) the 25 June visit; and (b) the 11 July visit. Correspondences are selected from both point clouds for rough alignment.

4.3. Point Cloud Registration To align the point clouds, we first scale the point cloud into a correct real-world unit in CloudCompare. To this end, we treat the point cloud in the second visit on 11 July as the reference point cloud. During this visit, three markers (M1, M2, and M3) were placed in the testbed as seen in Figure 12d. M1 and M2 were X marks made by the blue paint tape, while M3 was the UAV landing pad. The distances between the three markers were taken by a measuring tape (see the second column in Table 2). Next, the markers were visually identified from the dense point cloud (Figure 12a–c). The distances between three

Remote Sens. 2021, 13, 3152 markers in the point cloud were also measured (see the third column in Table 122). of Finally, 17 three scaling factors were calculated, based on which the average scaling factor of 1.054 m was applied for scaling the point cloud in the second visit.

Remote Sens. 2021, 13, x FOR PEER REVIEW 12 of 17

Table 2. Scaling factor calculation.

Marker Distance Field Measurement CloudCompare Measurement Scaling Factor Calculation M1 to M2 17.48 m 16.61 1.052 m/1 Average: 1.054 M1 to M3 10.06 m 9.52 1.057 m/1 m/1 FigureFigure 1M2. 2Deployment 12.to M3Deployment of markers of markers14.99 during m during the the visit visit on on 11 11July July where where14.22 ( a ()a )shows shows the the dense dense 1.054 point m/1 cloudcloud in in the the top top view; view; (b) shows(b the) shows blow the-up blow-up detail detailof the of point the point cloud cloud with with three three markers; markers; ( (cc)) shows blow-upblow-up details details of threeof three markers markers truncated truncated from fromthe dense the dense point point cloud; cloud; and and (d) ( dThereaftershows) shows the the, imagesthe images scaled of of markers point cloud collected collected in the by by first a smartphonea smartphonevisit was cameraaligned camera in to the thein ﬁeld. the reference field. point cloud through the registration protocol. Figure 11 demonstrates the selections of correspondences (A1-R1, A2-R2, A3-R3, and A4-R4) from both point clouds for rough align- Table 2. Scaling factor calculation. ment. Next, the point cloud in the first visit was further aligned by the automated ICP algorithm.Marker Field CloudCompare Scaling Factor Calculation DistanceFigure 13 showsMeasurement the comparison of Measurementpoint clouds under different views of the cliff fromM1 two to visits M2 during 17.48the registration m procedure. 16.61 The point cloud 1.052 from m/1 the 25 June visit Average: is renderedM1 to M3 in blue. As can 10.06 be m seen in Figure 13b,e,h, 9.52 small misalignments 1.057 m/1 can be observed 1.054 m/1 afterM rough 2 to M3 alignment. Such 14.99 mmisalignments can 14.22 be minimized after 1.054 fine m/1 alignment is performed (Figure 13c,f,i).

FigureFigure 13. 13.Investigation Investigation ofof pointpoint cloudcloud registration.registration. ( (a,,d,,g)) are are rough rough alignments alignments results results of of point point clouds clouds from from the the top, top, north, north, andand east east views; views; ( (bb,e,e,h,h)) areare thethe blow-upblow-up details;details; and ( c,,ff,,ii)) are are the the same same observation observation locations locations after after applying applying fine ﬁne alignment. alignment. The point cloud from the June 25 visit is rendered in blue. The point cloud from the June 25 visit is rendered in blue.

4.4.4.4. Cliff Monitoring The cloud-to-cloud distance is computed in CloudCompare and the results are The cloud-to-cloud distance is computed in CloudCompare and the results are shown shown in Figure 14. As can be seen from the figure, the majority of the cliff area is covered in Figure 14. As can be seen from the figure, the majority of the cliff area is covered in in green, indicating the discrepancies between two point clouds are about or less than 1.47 green, indicating the discrepancies between two point clouds are about or less than 1.47 cm cm (read from the figure). However, scattered yellow and red spots can be also found (read from the figure). However, scattered yellow and red spots can be also found from from the results. The cloud-to-cloud distances for these locations span from 19 cm (yellow) the results. The cloud-to-cloud distances for these locations span from 19 cm (yellow) to to 2.47 m (red), showing significant discrepancies that occurred in the point clouds from 2.47 m (red), showing significant discrepancies that occurred in the point clouds from two two field visits. field visits.

Remote Sens. 2021, 13, x FOR PEER REVIEW 13 of 17

Remote Sens. 2021, 13, 3152 13 of 17 Remote Sens. 2021, 13, x FOR PEER REVIEW 13 of 17

Figure 14. Cloud-to-cloud distance between point clouds from two visits from (a) bird’s-eye view; (b) top view; (c) east elevation view; and (d) north elevation view. Results are in log scale with the unit of m. FigureFigure 14.14.Cloud-to-cloud Cloud-to-cloud distancedistance betweenbetween pointpoint cloudsclouds fromfrom twotwo visitsvisits fromfrom ((aa)) bird’s-eyebird’s-eye view;view; ((bb)) toptop view;view; ((cc)) east east elevation elevation view;view; andand ((dd)) north north elevationelevation view.view. Results Results areare in in log log scale scale with with the the unit of m. To further investigateunit of m. such discrepancies, two locations (Patch A and B) are identified from the bird’s-eye view in Figure 14a. Patch A contains a steep cliff face covered by scat- ToTo further further investigate investigate such such discrepancies, discrepancies, two two locations locations (Patch (Patch A A and and B) B) areare identifiedidentified tered vegetation as shownfromfrom thethe in bird’s-eyebird’s Figure-eye 15view viewb; Patchin in Figure Figure B 14 locates14a.a. Patch Patch at A thecontains A contains flat atop steep a steepof cliff the cliff face cliff facecovered filled covered byby scat- by vegetation (Figure 15scatteredteredd). Asvegetation vegetationobserved as shown as from shown in theFigure in Figure figures, 15 b;15 Patchb; Patchthe B locatescloud B locates at-to the at-cloud the flat flat top distance top of ofthe the cliff cliffs filledare filled by large in the area of vegetationbyvegetation vegetation ( Figure(yellow (Figure 15 15d). spotsd). As As observed in observed Figure from from 15 thea the,c) figures, figures, and become the the cloud cloud-to-cloud smaller-to-cloud distancesdistancearounds areare largelarge in in the the area area of of vegetation vegetation (yellow (yellow spots spots in Figurein Figure 15a,c) 15a and,c) and become become smaller smaller around around the the cliff rock face (e.g.,cliffthe thecliff rock greenrock face (e.g.,face area ( thee.g., in green the Figure green area inarea15 Figurea). in ThisFigure 15a). is 15 This becausea). isThis because is thebecause the SfM SfM-MVS the-MVS SfM- algorithmMVSalgo- algorithm has difficultieshasrithm re difficultiesconstructing has difficulties reconstructing thin reconstructing structures thin structures thin such structures such as plants as plantssuch as[41 [41 plants],], leadingleading [41], toleading to reconstruction recon- to reconstruction errors to theerrorsstruction point to the clouds.errors point to clouds.the point clouds.

Figure 15. Investigation of Patch A and B in Figure 14 where (a,b) are cloud-to-cloud distance and the point cloud in the area defined by Patch A; and (c,d) are cloud-to-cloud distance and point cloud in the area defined by Patch B. FigureFigure 15. 15.Investigation Investigation of Patch of Patch A and BA in and Figure B 14in whereFigure (a, b14) are where cloud-to-cloud (a,b) are distance cloud- andto-cloud the point distance cloud in theand thearea point defined cloud by Patch in the A; and area (c, ddefined)To are reduce cloud-to-cloud by thePatch errors distanceA; andcaused and(c, dby point) arevegetation, cloud cloud in- theto we-cloud area truncate defined distance the by cloud Patch and B.-to point-cloud cloud distance in the area defined by Patchresult B.in Figure 14 by only reserving the steep cliff faces on the east and north sides. The newTo results reduce of cloud the errors-to-cloud caused distance by vegetation, are shown we in truncate Figure 16. the As cloud-to-cloud a result, the maximum distance result in Figure 14 by only reserving the steep cliff faces on the east and north sides. The To reduce the errorscloud- tocaused-cloud distanceby vegetation, has been reduced we truncate from 2.47 the m incloud Figure-to 14- cloudto 0.66 mdistance in Figure 16. newRed resultsspots can of be cloud-to-cloud still observed distance from the are figures, shown mainly in Figure caused 16. Asby the a result, scattered the maximumvegetation result in Figure 14 bycloud-to-cloudon only the cliff reserving faces. distance We thefurther has steep been inquiry cliff reduced three faces fromcloud on 2.47- tothe-cloud m east in Figuredistances and 14 north tofrom 0.66 typicalsides m in Figure. cliffThe faces 16. new results of cloud-Redandto-cloud spotsthe results can distance be range still observed from are 0.7shown fromcm to the in2.2 figures, Figurecm. Considering mainly 16. As caused athe result, size by theof thethe scattered entiremaximum vegetationcliff (about on30 them in cliff height), faces. such We further differences inquiry are three negligible cloud-to-cloud. distances from typical cliff faces cloud-to-cloud distanceand thehas results been rangereduced from 0.7from cm 2.47 to 2.2 m cm. in Considering Figure 14 the to size0.66 of m the in entire Figure cliff 16. (about Red spots can be still30 observed m in height), from such the differences figures, are mainly negligible. caused by the scattered vegetation on the cliff faces. We further inquiry three cloud-to-cloud distances from typical cliff faces and the results range from 0.7 cm to 2.2 cm. Considering the size of the entire cliff (about 30 m in height), such differences are negligible.

Remote Sens. 2021, 13, 3152 14 of 17 Remote Sens. 2021, 13, x FOR PEER REVIEW 14 of 17

FigureFigure 16. 16.Cloud-to-cloud Cloud-to-clouddistance distance and and the the point point cloud cloud of of the the truncated truncated cliff. cliff. ( a(a,b,b)) Results Results from from the the bird’s-eyebird’s-eye view;view; (c(c,d,d)) results results from from the the east east view; view; and and ( e(e,f,)f) results results from from the the north north view. view. The The unit unit in in (a,c,e) is m. (a,c,e) is m.

5.5. DiscussionDiscussions WeWe validated validated thethe proposed proposed method method through through a a few few small-scale small-scale experiments experiments using using a a rockrock sample. sample. AlthoughAlthough SfM-MVSSfM-MVS isis aa well-establishedwell-established workflow workflow for for reconstructing reconstructing point point clouds,clouds, few few studies studies in in the the literature literature focused focused on on the the robustness robustness of SfM-MVSof SfM-MVS against against different differ- lightingent lighting conditions conditions and and surface surface textures textures in thein the context context of of coastal coastal cliffs. cliffs. The The small-scale small-scale validationvalidation inin thisthis studystudyserves servesas as the the mean mean for for addressing addressing such such concerns. concerns. TheThe lighting lighting conditionsconditions andand surface textures textures (see (see the the second second and and third third columns columns in Table in Table 1) would1) would sim- simulateulate the the different different weather weather conditions conditions of a of cliff a cliff one one could could see see in the in the field. field. For For instance, instance, the thelighting lighting conditions conditions of ofthe the cliff cliff site site would would change change across across different different periods ofof thethe day;day;the the surfacesurface texture texture of of rock rock may may become become dark dark after after rain rain or or a typhoon. a typhoon. The The geometric geometric changes changes in small-scalein small-scale validation validation include include abrupt abrupt changes, changes such, such as adding as adding stones stones (Case (Case C); C); or gradualor grad- changes,ual changes such, such as adding as adding salt salt particle particle layers layers (Cases (Cases D D and and E). E). These These changes changes mimic mimic the the geomorphologicalgeomorphological changes of the the cliff. cliff. For For the the erosion erosion behavior behavior of the of thecliff, cliff, instead instead of add- of addinging contents, contents, landscape landscape features features of the of cliff the cliffwould would be removed. be removed. In this In case, this point case, clouds point cloudsin Case in C, Case D, and C, D, E andcan Ebe can considered be considered as the as initial the initial models, models, while while the point the point cloud cloud in Case in CaseA shall A shallbe the be new the newmodel model after aftererosion. erosion. ResultsResults from from the the small-scale small-scale validation validation demonstrated demonstrated the the effectiveness effectiveness of of our our method method inin detecting, detecting, identifying, identifying, and and quantifying quantifying geometric geometric changes changes in thein the rock rock sample, sample, regardless regard- ofless variations of variations in lighting in lighting conditions conditions and and surface surface texture. texture. Although Although the the cliff cliff in in the the field field validationvalidation of of this this study study did did not not experience experience visible visible erosion erosion due due to ato short a short inspection inspection interval, inter- theval, findings the findings in the in small-scale the small-scale validation validation would would serve serve as the as basis the basis for the for success the success of our of methodour method in monitoring in monitoring cliff erosioncliff erosion over over the long the term.long term. InIn termsterms of of correspondence correspondence selection, selection, fourfour pairs pairs of of correspondences correspondences are are selected selected on on thethe toptop ofof the the test test sample sample in in small-scale small-scale validationvalidation showingshowing inin FigureFigure5 .5. Selecting Selecting corre- corre- spondencesspondences from from other other locations locations of of the the rock rock sample sample is is also also feasible. feasible. Since Since correspondence correspondence selection only serves as the mean for rough alignment, errors that occurred in this regis- selection only serves as the mean for rough alignment, errors that occurred in this

Remote Sens. 2021, 13, 3152 15 of 17

tration stage can be further reduced during fine alignment and would not affect the final registration result. One difference between small-scale and field validations is that extra errors are in- duced in the field validation due to vegetation in the cliff area. Vegetation fully covers the top surface of the cliff and appears in scattered patterns at the vertical cliff faces. Estimating the locations of true rock surfaces in these areas from the point cloud could be very challenging as the surfaces are barely visible from UAV images. However, the false-positive results can be easily identified through visual inspections between cloud-to-cloud distance and ground truth measurements (see Figure 15). Since the nature of our method is a non-georeferenced approach, the point cloud generated by our method is not intended to contain any geographic information. Although most consumer-grade UAVs (including the ones in this study) provide geotagged images, such UAV images are not suitable for georeferencing due to the low accuracy of GIS coordinates. Secondly, the point cloud produced by our method cannot be directly linked to georeferenced datasets (e.g., geotagged maps, point clouds, or models). However, if a georeferenced point cloud of a cliff exists in the past, one can align a newly collected non-georeferenced point cloud from our method to the existing georeferenced one through the registration method established in this study. In terms of geomorphological changes, our method assumes that only a small portion of the cliff experiences erosion while the remainder of the cliff remains unchanged during inspections, which could be commonly found in coastal surveying [46]. Investigating dramatic geomorphological changes of a cliff due to severe erosions is out of the scope of this study.

6. Conclusions Monitoring cliff erosion is essential for maintaining a healthy coastal ecosystem. The usage of photogrammetry-based workflows and UAVs have been proven effective in monitoring coastal cliffs. To date, many photogrammetry-based methods rely on georeferencing frameworks for point cloud alignments. Despite the successes reported in these studies, georeferencing efforts significantly increase the project cost through securing high-end GPS equipment, hiring GIS specialists, and/or relying on GNSS-enabled UAVs. This may hinder the usage of photogrammetry technology for monitoring cliffs on a routine basis, particularly in underserved coastal communities where expensive hardware and trained GIS specialists are limited resources. In this study, we proposed a novel photogrammetry-based approach for identifying geomorphological changes of coastal cliffs that does not rely on any georeferencing efforts. The SfM-MVS algorithms were adopted in reconstructing 3D dense point clouds of the cliff. Then, a rigid registration protocol was established to gradually align two point clouds at different periods together to uncover the differential changes caused by cliff erosion. Our method has been examined by a series of small-scale experiments on a rock sample. Results indicated the proposed method can detect, localize, and quantify small changes that occurred in the rock sample, regardless of variations in lighting and surface texture conditions. Thereafter, we further validated our method on a full-scale coastal cliff in Guam. Point clouds from two field visits were reconstructed and aligned together to find the differential features caused by geomorphological changes. The findings of this study are highly impactful for being able to offer a low-cost and flexible cliff monitoring methodology to government agencies and stakeholders for their decision-making in coastal zone management.

Funding: This study is based on work supported by the seed grant through the National Science Foundation project Guam EPSCoR (Grant No. 1457769) in the United States. However, any opinions, ﬁndings, and conclusions, or recommendations expressed in this study are those of the author and do not necessarily reﬂect the views of the National Science Foundation or Guam EPSCoR. Data Availability Statement: The data presented in this study are available from the author upon reasonable request. Remote Sens. 2021, 13, 3152 16 of 17

Acknowledgments: This study was funded and the data was collected during the author’s previous term at the University of Guam. The author analyzed data and wrote the manuscript during the periods at both the University of Guam and Coastal Carolina University. The author wants to thank undergraduate student assistant Angela Maglaque for assisting in field visits; Maria Kottermair and Yuming Wen for providing fruitful discussions on equipment and software selections; the Guam EPSCoR program and the Research Corporation of the University of Guam for managing the grant; and the Office of Research and Sponsored Programs at the University of Guam for supporting a portion of the equipment used in this study. The author also wants to thank Colleen Bamba, Bastian Bentlage, Terry Donaldson, Roseann Jones, Jordan Jugo, and anonymous reviewers for improving the quality of this manuscript significantly. Conflicts of Interest: The author declares no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

References 1. Guam Coastal Zone Management Program and Draft Environmental Impact Statement; National Oceanic and Atmospheric Administra- tion, U.S. Department of Commerce: New York, NY, USA, 1979. 2. Emery, K.O.; Kuhn, G.G. Sea cliffs: Their processes, profiles, and classification. Geol. Soc. Am. Bull. 1982, 93, 644–654. [CrossRef] 3. Guam Hazard Mitigation Plan, Guam Homeland Security Office. 2019. Available online: https://ghs.guam.gov/sites/default/ files/final_2019_guam_hmp_20190726.pdf (accessed on 7 August 2021). 4. Thieler, E.R.; Danforth, W.W. Historical shoreline mapping (II): Application of the digital shoreline mapping and analysis systems (DSMS/DSAS) to shoreline change mapping in Puerto Rico. J. Coast. Res. 1994, 10, 600–620. 5. Sear, D.A.; Bacon, S.R.; Murdock, A.; Doneghan, G.; Baggaley, P.; Serra, C.; LeBas, T.P. Cartographic, geophysical and diver surveys of the medieval town site at Dunwich, Suffolk, England. Int. J. Naut. Archaeol. 2011, 40, 113–132. [CrossRef] 6. Westoby, M.J.; Lim, M.; Hogg, M.; Pound, M.J.; Dunlop, L.; Woodward, J. Cost-effective erosion monitoring of coastal cliffs. Coast. Eng. 2018, 138, 152–164. [CrossRef] 7. Letortu, P.; Jaud, M.; Grandjean, P.; Ammann, J.; Costa, S.; Maquaire, O.; Delacourt, C. Examining high-resolution survey methods for monitoring cliff erosion at an operational scale. GIScience Remote Sens. 2018, 55, 457–476. [CrossRef] 8. Rosser, N.J.; Petley, D.N.; Lim, M.; Dunning, S.A.; Allison, R.J. Terrestrial laser scanning for monitoring the process of hard rock coastal cliff erosion. Q. J. Eng. Geol. Hydrogeol. 2005, 38, 363–375. [CrossRef] 9. Hayakawa, Y.S.; Obanawa, H. Volumetric Change Detection in Bedrock Coastal Cliffs Using Terrestrial Laser Scanning and UAS-Based SfM. Sensors 2020, 20, 3403. [CrossRef] 10. Klemas, V.V. Coastal and environmental remote sensing from unmanned aerial vehicles: An overview. J. Coast. Res. 2015, 31, 1260–1267. [CrossRef] 11. Crommelinck, S.; Bennett, R.; Gerke, M.; Nex, F.; Yang, M.Y.; Vosselman, G. Review of automatic feature extraction from high-resolution optical sensor data for UAV-based cadastral mapping. Remote Sens. 2016, 8, 689. [CrossRef] 12. Nex, F.; Remondino, F. UAV for 3D mapping applications: A review. Appl. Geomat. 2014, 6, 1–15. [CrossRef] 13. Duffy, J.P.; Shutler, J.D.; Witt, M.J.; DeBell, L.; Anderson, K. Tracking fine-scale structural changes in coastal dune morphology using kite aerial photography and uncertainty-assessed structure-from-motion photogrammetry. Remote Sens. 2018, 10, 1494. [CrossRef] 14. Fonstad, M.A.; Dietrich, J.T.; Courville, B.C.; Jensen, J.L.; Carbonneau, P.E. Topographic structure from motion: A new development in photogrammetric measurement. Earth Surf. Process. Landf. 2013, 38, 421–430. [CrossRef] 15. Jaud, M.; Bertin, S.; Beauverger, M.; Augereau, E.; Delacourt, C. RTK GNSS-Assisted Terrestrial SfM Photogrammetry without GCP: Application to Coastal Morphodynamics Monitoring. Remote Sens. 2020, 12, 1889. [CrossRef] 16. Forlani, G.; Pinto, L.; Roncella, R.; Pagliari, D. Terrestrial photogrammetry without ground control points. Earth Sci. Inform. 2014, 7, 71–81. [CrossRef] 17. Chiang, K.W.; Tsai, M.L.; Chu, C.H. The development of an UAV borne direct georeferenced photogrammetric platform for ground control point free applications. Sensors 2012, 12, 9161–9180. [CrossRef][PubMed] 18. Turner, D.; Lucieer, A.; Wallace, L. Direct georeferencing of ultrahigh-resolution UAV imagery. IEEE Trans. Geosci. Remote Sens. 2013, 52, 2738–2745. [CrossRef] 19. Taddia, Y.; Stecchi, F.; Pellegrinelli, A. Coastal Mapping using DJI Phantom 4 RTK in Post-Processing Kinematic Mode. Drones 2020, 4, 9. [CrossRef] 20. Peppa, M.V.; Hall, J.; Goodyear, J.; Mills, J.P. Photogrammetric assessment and comparison of DJI Phantom 4 pro and phantom 4 RTK small unmanned aircraft systems. ISPRS Geospat. Week 2019, XLII-2/W13, 503–509. [CrossRef] 21. Urban, R.; Reindl, T.; Brouˇcek,J. Testing of drone DJI Phantom 4 RTK accuracy. In Advances and Trends in Geodesy, Cartography and Geoinformatics II. In Proceedings of the 11th International Scientific and Professional Conference on Geodesy, Cartography and Geoinformatics (GCG 2019), Demänovská Dolina, Low Tatras, Slovakia, 10–13 September 2019; CRC Press: Boca Raton, FA, USA; p. 99. Remote Sens. 2021, 13, 3152 17 of 17

22. Chen, Y.; Medioni, G. Object modelling by registration of multiple range images. Image Vis. Comput. 1992, 10, 145–155. [CrossRef] 23. Besl, P.J.; McKay, N.D. Method for registration of 3-D shapes. In Sensor Fusion IV: Control Paradigms and Data Structures; International Society for Optics and Photonics: Washington, DC, USA, 1992; Volume 1611, pp. 586–606. 24. Costantino, D.; Settembrini, F.; Pepe, M.; Alﬁo, V.S. Develop of New Tools for 4D Monitoring: Case Study of Cliff in Apulia Region (Italy). Remote Sens. 2021, 13, 1857. [CrossRef] 25. Obanawa, H.; Hayakawa, Y.S. Variations in volumetric erosion rates of bedrock cliffs on a small inaccessible coastal island determined using measurements by an unmanned aerial vehicle with structure-from-motion and terrestrial laser scanning. Prog. Earth Planet. Sci. 2018, 5, 1–10. [CrossRef] 26. Michoud, C.; Carrea, D.; Costa, S.; Derron, M.H.; Jaboyedoff, M.; Delacourt, C.; Davidson, R. Landslide detection and monitoring capability of boat-based mobile laser scanning along Dieppe coastal cliffs, Normandy. Landslides 2015, 12, 403–418. [CrossRef] 27. Dewez, T.J.B.; Leroux, J.; Morelli, S. Cliff Collapse Hazard from Repeated Multicopter UAV Acquisitions: Return on Experience. International Archives of the Photogrammetry. Remote Sens. Spat. Inf. Sci. 2016, 41, 805–811. [CrossRef] 28. Stumpf, A.; Malet, J.P.; Allemand, P.; Pierrot-Deseilligny, M.; Skupinski, G. Ground-based multi-view photogrammetry for the monitoring of landslide deformation and erosion. Geomorphology 2015, 231, 130–145. 29. Long, N.; Millescamps, B.; Pouget, F.; Dumon, A.; Lachaussee, N.; Bertin, X. Accuracy assessment of coastal topography derived from UAV images. The International Archives of Photogrammetry. Remote Sens. Spat. Inf. Sci. 2016, 41, 1127. 30. Chaiyasarn, K.; Kim, T.K.; Viola, F.; Cipolla, R.; Soga, K. Distortion-free image mosaicing for tunnel inspection based on robust cylindrical surface estimation through structure from motion. J. Comput. Civ. Eng. 2016, 30, 04015045. [CrossRef] 31. Dietrich, J.T. Bathymetric structure–from–motion: Extracting shallow stream bathymetry from multi–view stereo photogrammetry. Earth Surf. Process. Landf. 2017, 42, 355–364. [CrossRef] 32. Bhadrakom, B.; Chaiyasarn, K. As-built 3D modeling based on structure from motion for deformation assessment of historical buildings. Int. J. Geomate 2016, 11, 2378–2384. [CrossRef] 33. Lowe, D.G. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis. 2004, 60, 91–110. [CrossRef] 34. Shi, J. Good features to track. In Proceedings of the Computer Vision and Pattern Recognition, Proceedings CVPR’94, IEEE Computer Society Conference on IEEE, Seattle, WA, USA, 21–23 June 1994; pp. 593–600. 35. Rosten, E.; Drummond, T. Fusing points and lines for high performance tracking. In Proceedings of the Tenth IEEE International Conference on Computer Vision (ICCV’05), Beijing, China, 17–21 October 2005; Volumes 1–2, pp. 1508–1515. 36. Harris, C.; Stephens, M. A combined corner and edge detector. Alvey Vis. Conf. 1988, 15, 50. 37. Leutenegger, S.; Chli, M.; Siegwart, R.Y. BRISK: Binary robust invariant scalable keypoints. In Proceedings of the 2011 International Conference on Computer Vision, Barcelona, Spain, 6–13 November 2011; pp. 2548–2555. 38. Bay, H.; Tuytelaars, T.; Van Gool, L. Surf: Speeded up robust features. In European Conference on Computer Vision; Springer: Berlin/Heidelberg, Germany, 2006; pp. 404–417. 39. James, M.R.; Robson, S. Straightforward reconstruction of 3D surfaces and topography with a camera: Accuracy and geoscience application. J. Geophys. Res. Earth Surf. 2012, 117, F03017. [CrossRef] 40. Furukawa, Y.; Hernández, C. Multi-view stereo: A tutorial. Found. Trends Comput. Graph. Vis. 2015, 9, 1–148. [CrossRef] 41. Smith, M.W.; Carrivick, J.L.; Quincey, D.J. Structure from motion photogrammetry in physical geography. Prog. Phys. Geogr. 2016, 40, 247–275. [CrossRef] 42. AgiSoft Metashape Professional (Version 1.6.2) (Software). 2020. Available online: http://www.agisoft.com/downloads/installer/ (accessed on 1 April 2020). 43. CloudCompare (version 2.10.2) [GPL Software]. 2020. Available online: http://www.cloudcompare.org/ (accessed on 1 June 2020). 44. Pix4Dcapture (Version 4.10.0) (Smartphone app). 2020. Available online: https://www.pix4d.com/product/pix4dcapture (accessed on 1 June 2020). 45. DJI GO 4 (Version 4.3.36) (Smartphone app). 2020. Available online: https://www.dji.com/downloads/djiapp/dji-go-4 (accessed on 1 June 2020). 46. Grottoli, E.; Biausque, M.; Rogers, D.; Jackson, D.W.; Cooper, J.A.G. Structure-from-Motion-Derived Digital Surface Models from Historical Aerial Photographs: A New 3D Application for Coastal Dune Monitoring. Remote Sens. 2021, 13, 95. [CrossRef]