Datasets:

---

LLMDH

/

OpenScience

like 0

Follow

LLM - Digital Humanities 16

ERROR: type should be string, got "https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Beyond 2D inventories : synoptic 3D landslide volume calculation \nfrom repeat LiDAR data Thomas G. Bernard, Dimitri Lague, Philippe Steer \nUniv Rennes, CNRS, Géosciences Rennes - UMR 6118, 35000, Rennes, France 5 Abstract. Efficient and robust landslide mapping and volume estimation is essential to rapidly infer landslide spatial distribution, to \nquantify the role of triggering events on landscape changes and to assess direct and secondary landslide-related geomorphic \nhazards. Many efforts have been made during the last decades to develop landslide areal mapping methods, based on 2D quantify the role of triggering events on landscape changes and to assess direct and secondary landslide-related geomorphic \nhazards. Many efforts have been made during the last decades to develop landslide areal mapping methods, based on 2D \nsatellite or aerial images, and to constrain empirical volume-area (V-A) allowing in turn to offer indirect estimates of landslide \n10 \nvolume. Despite these efforts, some major issues remain including the uncertainty of the V-A scaling, landslide amalgamation \nand the under-detection of reactivated landslides. To address these issues, we propose a new semi-automatic 3D point cloud \ndifferencing method to detect geomorphic changes, obtain robust landslide inventories and directly measure the volume and \ngeometric properties of landslides. This method is based on the M3C2 algorithm and was applied to a multi-temporal airborne \nLiDAR dataset of the Kaikoura region New Zealand following the M 7 8 earthquake of 14 November 2016\n15 satellite or aerial images, and to constrain empirical volume-area (V-A) allowing in turn to offer indirect estimates of landslide \n10 \nvolume. Despite these efforts, some major issues remain including the uncertainty of the V-A scaling, landslide amalgamation \nand the under-detection of reactivated landslides. To address these issues, we propose a new semi-automatic 3D point cloud \ndifferencing method to detect geomorphic changes, obtain robust landslide inventories and directly measure the volume and \ngeometric properties of landslides. This method is based on the M3C2 algorithm and was applied to a multi-temporal airborne LiDAR dataset of the Kaikoura region, New Zealand, following the Mw 7.8 earthquake of 14 November 2016. 15 \nWe demonstrate that 3D point cloud differencing offers a greater sensitivity to detect small changes than a classical difference \nof DEMs (digital elevation models). In a small 5 km² area, prone to landslide reactivation and amalgamation, where a previous \nstudy identified 27 landslides, our method is able to detect 1431 landslide sources and 853 deposits with a total volume of \n908,055 ± 215,640 m3 and 1,008,626 ± 172,745 m3, respectively. Abstract. This high number of landslides is set by the ability of our LiDAR dataset of the Kaikoura region, New Zealand, following the Mw 7.8 earthquake of 14 November 2016. 15 \nWe demonstrate that 3D point cloud differencing offers a greater sensitivity to detect small changes than a classical difference \nof DEMs (digital elevation models). In a small 5 km² area, prone to landslide reactivation and amalgamation, where a previous \nstudy identified 27 landslides, our method is able to detect 1431 landslide sources and 853 deposits with a total volume of \n908,055 ± 215,640 m3 and 1,008,626 ± 172,745 m3, respectively. This high number of landslides is set by the ability of our method to detect subtle changes and therefore small landslides with a carefully constrained lower limit of 20 m² (90% with \n20 \nA<300 m²). Moreover, the analysis of landslide geometric properties reveals the absence of a rollover in the landslide area \ndistribution, which is a feature classically described in the literature. This result suggests that the rollover behaviour previously \nobserved is due to an under detection of small landslides. Reactivated landslides represent 27.2 % of the total landslide source \narea and 29.9 ± 12.8 % of the total volume. Reactivated landslides are located in areas where landslide mapping methods based method to detect subtle changes and therefore small landslides with a carefully constrained lower limit of 20 m² (90% with \n20 \nA<300 m²). Moreover, the analysis of landslide geometric properties reveals the absence of a rollover in the landslide area \ndistribution, which is a feature classically described in the literature. This result suggests that the rollover behaviour previously \nobserved is due to an under detection of small landslides. Reactivated landslides represent 27.2 % of the total landslide source \narea and 29.9 ± 12.8 % of the total volume. Reactivated landslides are located in areas where landslide mapping methods based on 2D images are assumed to perform poorly due to the weak contrast in texture and colour between the two epochs. Our result \n25 \ntherefore suggests that the number, area and volume of landslides can be significantly under-estimated by these methods. To \nour knowledge, this is the first approach to create a regional landslide inventory map from 3D point cloud differencing. Abstract. Our \nresults call for a more systematic use of high-resolution 3D topographic data to assess the impact of extreme events on \ntopographic changes in regions prone to landsliding and to infer the geometric scaling properties of landslides. 30 on 2D images are assumed to perform poorly due to the weak contrast in texture and colour between the two epochs. Our result \n25 \ntherefore suggests that the number, area and volume of landslides can be significantly under-estimated by these methods. To \nour knowledge, this is the first approach to create a regional landslide inventory map from 3D point cloud differencing. Our \nresults call for a more systematic use of high-resolution 3D topographic data to assess the impact of extreme events on \ntopographic changes in regions prone to landsliding and to infer the geometric scaling properties of landslides. 30 1 1 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 1. Introduction This uncertainty could lead to an order magnitude of difference in total estimated volume given the non-linearity of eq. (1) (Larsen et al., 2010). Two other sources of error arise from the detectability of \n45 \nindividual landslides themselves and the ability to accurately measure the distribution of landslide areas, due to landslide \namalgamation and under-detection of reactivated landslides. Landslide amalgamation can produce up to 200 % error in the \ntotal volume estimation (Li et al., 2014; Marc and Hovius, 2015) and occurs because of landslide spatial clustering or incorrect \nmapping due, for instance, to automatic processing. Indeed, automatic landslide mapping (Behling et al., 2014; Marc et al., volume given the non-linearity of eq. (1) (Larsen et al., 2010). Two other sources of error arise from the detectability of \n45 \nindividual landslides themselves and the ability to accurately measure the distribution of landslide areas, due to landslide \namalgamation and under-detection of reactivated landslides. Landslide amalgamation can produce up to 200 % error in the \ntotal volume estimation (Li et al., 2014; Marc and Hovius, 2015) and occurs because of landslide spatial clustering or incorrect \nmapping due, for instance, to automatic processing. Indeed, automatic landslide mapping (Behling et al., 2014; Marc et al., 2019; Martha et al., 2010; Pradhan et al., 2016) relies on the difference in texture, color and spectral properties such as NDVI \n50 \n(Normalized difference vegetation index) of multispectral 2D images between pre- and post-landslide images, assuming that \nlandslides lead to vegetation removal or significant texture change. During this process, difficulties in automatic segmentation \nof landslide sources can result in incorrect estimate of individual landslide area, which propagates into a much larger estimate \nof volume owing to the non-linearity of eq. (1). Manual mapping and automatic algorithms based on geometrical and 2019; Martha et al., 2010; Pradhan et al., 2016) relies on the difference in texture, color and spectral properties such as NDVI \n50 \n(Normalized difference vegetation index) of multispectral 2D images between pre- and post-landslide images, assuming that \nlandslides lead to vegetation removal or significant texture change. During this process, difficulties in automatic segmentation \nof landslide sources can result in incorrect estimate of individual landslide area, which propagates into a much larger estimate \nof volume owing to the non-linearity of eq. (1). 1. Introduction Manual mapping and automatic algorithms based on geometrical and topographical inconsistencies can enable to reduce the amalgamation effect on landslide volume estimation (Marc and Hovius, \n55 \n2015), but it remains a source of error due to the inherent spatial clustering of landslides and the overlapping of landslide \ndeposits and sources. Under-detection of reactivated landslides occurs because the spectral signature of images is not \nsufficiently altered by a new failure. This phenomenon is particularly important in areas with thin soils and sparse or total lack \nof vegetation (Behling et al., 2014). Guzetti et al. (2009) show that the proportion of landslide volume mobilized by topographical inconsistencies can enable to reduce the amalgamation effect on landslide volume estimation (Marc and Hovius, \n55 \n2015), but it remains a source of error due to the inherent spatial clustering of landslides and the overlapping of landslide \ndeposits and sources. Under-detection of reactivated landslides occurs because the spectral signature of images is not \nsufficiently altered by a new failure. This phenomenon is particularly important in areas with thin soils and sparse or total lack \nof vegetation (Behling et al., 2014). Guzetti et al. (2009) show that the proportion of landslide volume mobilized by topographical inconsistencies can enable to reduce the amalgamation effect on landslide volume estimation (Marc and Hovius, \n55 \n2015), but it remains a source of error due to the inherent spatial clustering of landslides and the overlapping of landslide \ndeposits and sources. Under-detection of reactivated landslides occurs because the spectral signature of images is not \nsufficiently altered by a new failure. This phenomenon is particularly important in areas with thin soils and sparse or total lack \nof vegetation (Behling et al., 2014). Guzetti et al. (2009) show that the proportion of landslide volume mobilized by reactivations can reach 62% for an individual snowmelt or rainfall events. Yet, the level of landslide reactivation in a given \n60 \ninventory remains generally largely unknown. The delimitation of reactivated landslides is therefore critical to robustly infer \ntotal landslide volume. To better detect reactivated landslides, Behling et al. (2014, 2016) developed a method using temporal reactivations can reach 62% for an individual snowmelt or rainfall events. Yet, the level of landslide reactivation in a given \n60 \ninventory remains generally largely unknown. The delimitation of reactivated landslides is therefore critical to robustly infer \ntotal landslide volume. To better detect reactivated landslides, Behling et al. 1. Introduction In mountain areas, extreme events such as large earthquakes and typhoons can trigger thousands of landslides. Landslides are \na key agent of hillslope and landscape erosion (Keefer, 1994; Malamud et al., 2004) and represent a significant hazard for local \npopulations. Efficient and rapid mapping of landslides is required to robustly infer their spatial distribution, their total volume, \nthe induced landscape changes and the associated direct and secondary hazards (Guzzetti et al., 2012; Hovius et al., 1997; \n35 \nMarc et al., 2016; Parker et al., 2011). Following a triggering event, total landslide volume over a regional scale is classically \ndetermined in two steps: (i) individual landslide mapping using 2D satellite or aerial images (e.g., Behling et al., 2014; Fan et \nal, 2019; Guzetti et al., 2012; Li et al., 2014; Martha et al., 2010; Massey et al., 2018; Parker et al. 2011) and (ii) indirect \nvolume estimation using a volume-area relationship (e.g. Simonett, 1967; Larsen et al., 2010): g\np\ng\n𝑉= 𝛼𝐴𝛾\n(1) \n40 \nwith V and A the volume and area of individual landslides, 𝛼 a prefactor and 𝛾 a scaling exponent ranging between 1.3 and 1.6 \n(e.g., Larsen et al., 2010, Massey et al., 2020). A first source of error comes from the uncertainty on the values of 𝛼 and 𝛾 which tend to be site specific and process specific \n(e.g. shallow versus bedrock landsliding). This uncertainty could lead to an order magnitude of difference in total estimated 𝑉= 𝛼𝐴𝛾\n(1) \n40 \nwith V and A the volume and area of individual landslides, 𝛼 a prefactor and 𝛾 a scaling exponent ranging between 1.3 and 1.6 \n(e.g., Larsen et al., 2010, Massey et al., 2020). A first source of error comes from the uncertainty on the values of 𝛼 and 𝛾 which tend to be site specific and process specific \n(e.g. shallow versus bedrock landsliding). This uncertainty could lead to an order magnitude of difference in total estimated 𝑉= 𝛼𝐴𝛾\n(1) 40 𝑉= 𝛼𝐴𝛾 (1) with V and A the volume and area of individual landslides, 𝛼 a prefactor and 𝛾 a scaling exponent ranging between 1.3 and 1.6 \n(e.g., Larsen et al., 2010, Massey et al., 2020). (e.g., Larsen et al., 2010, Massey et al., 2020). A first source of error comes from the uncertainty on the values of 𝛼 and 𝛾 which tend to be site specific and process specific \n(e.g. shallow versus bedrock landsliding). 1. Introduction Even though this method is fast and works properly on planar surface, a vertical difference can \nbe prone to strong errors when used to quantify changes on vertical or very steep surfaces where landsliding typically occurs \n(e.g., Lague et al., 2013). The “Multiscale model-to-model cloud comparison” (M3C2) algorithm implemented by Lague et al. 2015; Ventura et al., 2011). The most commonly used technique is the difference of DEM (DoD) which consists in computing \n70 \nthe vertical elevation differences between two DEMs of different time (Corsini et al., 2009; Giordan et al., 2013; Mora et al., \n2018; Wheaton et al., 2010). Even though this method is fast and works properly on planar surface, a vertical difference can \nbe prone to strong errors when used to quantify changes on vertical or very steep surfaces where landsliding typically occurs \n(e.g., Lague et al., 2013). The “Multiscale model-to-model cloud comparison” (M3C2) algorithm implemented by Lague et al. (2013) rather considers a direct 3D point cloud comparison. This algorithm has three main advantages over a DoD: (i) it \n75 \noperates directly on 3D point clouds, avoiding a phase of DEM creation that is conducive to a loss of resolution imposed by \nthe pixel size and potential data interpolation, (ii) it computes 3D distances along the normal direction of the topographic \nsurface, allowing to better capture subtle changes on steep surfaces, (iii) it computes a spatially variable confidence interval \nthat accounts for surface roughness, point density and uncertainties in data registration. Applicable to any type of 3D data to measure the orthogonal distance between two point clouds, this approach has generally been used for Terrestrial Lidar and \n80 \nUAV photogrammetry over sub-kilometer scales. In the context of landsliding, it has been used to infer the displacement and \nvolume of individual landslides, using point clouds obtained by UAV photogrammetry (e.g., Esposito et al., 2017; Stumpf et \nal., 2015), as well as for rockfall studies (Benjamin et al.2016; Williams et al., 2018). Yet, to our knowledge, the systematic \ndetection and segmentation of hundreds of landslides, at a regional scale, from 3D point cloud have not yet been attempted. 80 measure the orthogonal distance between two point clouds, this approach has generally been used for Terrestrial Lidar and \n80 \nUAV photogrammetry over sub-kilometer scales. 1. Introduction Even though this method is fast and works properly on planar surface, a vertical difference can \nbe prone to strong errors when used to quantify changes on vertical or very steep surfaces where landsliding typically occurs \n(e.g., Lague et al., 2013). The “Multiscale model-to-model cloud comparison” (M3C2) algorithm implemented by Lague et al. (2013) rather considers a direct 3D point cloud comparison. This algorithm has three main advantages over a DoD: (i) it \n75 \noperates directly on 3D point clouds, avoiding a phase of DEM creation that is conducive to a loss of resolution imposed by \nthe pixel size and potential data interpolation, (ii) it computes 3D distances along the normal direction of the topographic \nsurface, allowing to better capture subtle changes on steep surfaces, (iii) it computes a spatially variable confidence interval \nthat accounts for surface roughness, point density and uncertainties in data registration. Applicable to any type of 3D data to \nmeasure the orthogonal distance between two point clouds, this approach has generally been used for Terrestrial Lidar and \n80 \nUAV photogrammetry over sub-kilometer scales. In the context of landsliding, it has been used to infer the displacement and \nvolume of individual landslides, using point clouds obtained by UAV photogrammetry (e.g., Esposito et al., 2017; Stumpf et \nal., 2015), as well as for rockfall studies (Benjamin et al.2016; Williams et al., 2018). Yet, to our knowledge, the systematic \ndetection and segmentation of hundreds of landslides, at a regional scale, from 3D point cloud have not yet been attempted. 2015; Ventura et al., 2011). The most commonly used technique is the difference of DEM (DoD) which consists in computing \n70 \nthe vertical elevation differences between two DEMs of different time (Corsini et al., 2009; Giordan et al., 2013; Mora et al., \n2018; Wheaton et al., 2010). Even though this method is fast and works properly on planar surface, a vertical difference can \nbe prone to strong errors when used to quantify changes on vertical or very steep surfaces where landsliding typically occurs \n(e.g., Lague et al., 2013). The “Multiscale model-to-model cloud comparison” (M3C2) algorithm implemented by Lague et al. 2015; Ventura et al., 2011). The most commonly used technique is the difference of DEM (DoD) which consists in computing \n70 \nthe vertical elevation differences between two DEMs of different time (Corsini et al., 2009; Giordan et al., 2013; Mora et al., \n2018; Wheaton et al., 2010). 1. Introduction (2014, 2016) developed a method using temporal 2 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. NDVI-trajectories which describe temporal footprints of vegetation changes but cannot fully address complex cases when \ntexture is not significantly changing such as bedrock landsliding on bare rock hillslopes. Addressing these three sources of uncertainty - volume-area scaling uncertainty, landslide amalgamation and the under-\n65 \ndetection of landslide reactivation - requires to explore new approaches to obtain and analyse landslide inventories. In the last \ndecade, the increasing availability of multi-temporal high resolution 3D point cloud data and digital elevation models (DEM), \nbased on aerial or satellite photogrammetry and Light Detection and Ranging (LiDAR), has opened the possibility to better \nquantify landside volume and displacement (Bull et al., 2010; Mouyen et al., 2019; Okyay et al., 2019; Passalacqua et al., Addressing these three sources of uncertainty - volume-area scaling uncertainty, landslide amalgamation and the under-\n65 \ndetection of landslide reactivation - requires to explore new approaches to obtain and analyse landslide inventories. In the last \ndecade, the increasing availability of multi-temporal high resolution 3D point cloud data and digital elevation models (DEM), \nbased on aerial or satellite photogrammetry and Light Detection and Ranging (LiDAR), has opened the possibility to better \nquantify landside volume and displacement (Bull et al., 2010; Mouyen et al., 2019; Okyay et al., 2019; Passalacqua et al., 65 Addressing these three sources of uncertainty - volume-area scaling uncertainty, landslide amalgamation and the under-\n65 \ndetection of landslide reactivation - requires to explore new approaches to obtain and analyse landslide inventories. In the last \ndecade, the increasing availability of multi-temporal high resolution 3D point cloud data and digital elevation models (DEM), \nbased on aerial or satellite photogrammetry and Light Detection and Ranging (LiDAR), has opened the possibility to better \nquantify landside volume and displacement (Bull et al., 2010; Mouyen et al., 2019; Okyay et al., 2019; Passalacqua et al., \n2015; Ventura et al., 2011). The most commonly used technique is the difference of DEM (DoD) which consists in computing \n70 \nthe vertical elevation differences between two DEMs of different time (Corsini et al., 2009; Giordan et al., 2013; Mora et al., \n2018; Wheaton et al., 2010). 1. Introduction In the context of landsliding, it has been used to infer the displacement and \nvolume of individual landslides, using point clouds obtained by UAV photogrammetry (e.g., Esposito et al., 2017; Stumpf et \nal., 2015), as well as for rockfall studies (Benjamin et al.2016; Williams et al., 2018). Yet, to our knowledge, the systematic \ndetection and segmentation of hundreds of landslides, at a regional scale, from 3D point cloud have not yet been attempted. Here, we produce an inventory map of landslide topographic changes using a semi-automatic 3D point cloud differencing \n85 \nmethod based on M3C2 and applied to multi-temporal airborne LiDAR data. We investigate the potential of this method to \nrobustly infer landslide volumes in a region prone to landslide reactivation and amalgamation. We applied our method to a \ncomplex topography located near Kaikoura, New Zealand, where the 2016 Mw 7.8 earthquake triggered nearly 30,000 \nlandslides over a 10,000 km² area (Massey et al., 2020). We choose a 5 km² area characterized by a high landslide spatial Here, we produce an inventory map of landslide topographic changes using a semi-automatic 3D point cloud differencing \n85 \nmethod based on M3C2 and applied to multi-temporal airborne LiDAR data. We investigate the potential of this method to \nrobustly infer landslide volumes in a region prone to landslide reactivation and amalgamation. We applied our method to a \ncomplex topography located near Kaikoura, New Zealand, where the 2016 Mw 7.8 earthquake triggered nearly 30,000 \nlandslides over a 10,000 km² area (Massey et al., 2020). We choose a 5 km² area characterized by a high landslide spatial density along the Conway segment of the Hope fault, inactive during the earthquake, where pre- and post-earthquake LiDAR \n90 \nwere available and investigate the proportion of reactivated landslide volume on the total budget (Fig. 1). This area has a \nvariety of vegetation cover (e.g. trees, shrubs, bare bedrock) and pre-existing landslides, and typically represents a challenge \nfor conventional 2D landslide mapping. We illustrate the benefits of working directly on 3D data to generate landslide \ninventories, and discuss the methodological advantages to operate directly on point clouds with M3C2 compared to DoD in terms of detection accuracy and error for total landslide volume. 95 3 3 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 1. Introduction The paper is organized as followed: first, the LiDAR dataset is presented followed by a description of the 3D point cloud \ndifferencing method. Second, results of the geomorphic changes detection and landslides individualization method applied to \nthe study area are presented. Detected landslides properties are then investigated in terms of area and volume, specifically for \nreactivated and new landslides. Finally, current limitations of the method are discussed as well as knowledge gained on the \nimportance of landslide reactivation on co-seismic landslide inventory budget. 100 Figure 1: Maps of the regional context, location and visualization of the study area. (a) Regional map of Kaikoura with the location \nof the 2016 Mw 7.8 earthquake, associated active faults and the study area. (b) Orthophotos focused on the study area dated before \nand after the earthquake with the 5 km² LiDAR dataset extent used in this paper (all imageries are available at \nhttps://data.linz.govt.nz/set/4702-nz-aerial-imagery/, Aerial survey 2017). Figure 1: Maps of the regional context, location and visualization of the study area. (a) Regional map of Kaikoura with the location \nof the 2016 Mw 7.8 earthquake, associated active faults and the study area. (b) Orthophotos focused on the study area dated before \nand after the earthquake with the 5 km² LiDAR dataset extent used in this paper (all imageries are available at \nhttps://data.linz.govt.nz/set/4702-nz-aerial-imagery/, Aerial survey 2017). 2. Data description In this study, we compare two 3D point clouds obtained from airborne LiDAR data before and after the November 14 2016 \nKaikoura earthquake (Table. 1). Both airborne LiDAR surveys were acquired during summer. Pre-earthquake LiDAR data \nrepresents a combination of six flights performed from March 13, 2014 to March 20, 2014 for a resulting ground point density In this study, we compare two 3D point clouds obtained from airborne LiDAR data before and after the November 14 2016 \nKaikoura earthquake (Table. 1). Both airborne LiDAR surveys were acquired during summer. Pre-earthquake LiDAR data \nrepresents a combination of six flights performed from March 13, 2014 to March 20, 2014 for a resulting ground point density \nof 3.8 ± 2.1 pts/m². The vertical accuracy of this dataset has been estimated at 0.068 m to 0.165 m from check points located \n110 \non highways (Dolan and Rhodes, 2016). The post-earthquake LiDAR survey took place rapidly after the earthquake from \nDecember 3, 2016 to January 6, 2017 for an average ground point density of 11.5 ± 6.8 pts/m². Vertical accuracy of this dataset In this study, we compare two 3D point clouds obtained from airborne LiDAR data before and after the November 14 2016 \nKaikoura earthquake (Table. 1). Both airborne LiDAR surveys were acquired during summer. Pre-earthquake LiDAR data \nrepresents a combination of six flights performed from March 13, 2014 to March 20, 2014 for a resulting ground point density \nof 3.8 ± 2.1 pts/m². The vertical accuracy of this dataset has been estimated at 0.068 m to 0.165 m from check points located \n110 \non highways (Dolan and Rhodes, 2016). The post-earthquake LiDAR survey took place rapidly after the earthquake from \nDecember 3, 2016 to January 6, 2017 for an average ground point density of 11.5 ± 6.8 pts/m². Vertical accuracy of this dataset of 3.8 ± 2.1 pts/m². The vertical accuracy of this dataset has been estimated at 0.068 m to 0.165 m from check points located \n110 \non highways (Dolan and Rhodes, 2016). The post-earthquake LiDAR survey took place rapidly after the earthquake from \nDecember 3, 2016 to January 6, 2017 for an average ground point density of 11.5 ± 6.8 pts/m². Vertical accuracy of this dataset 4 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. assessed on ground classified points is 0.04 m (Aerial survey, 2017). 2. Data description The difference in acquisition dates represents a period of \n2 years and 8 months. The vegetation of both datasets was removed using the classification provided by the downloaded data \nto keep only ground points. In addition to LiDAR data, orthophotos were used to visually validate the detection of landslides \n115 \nfrom the 3D approach for new landslide sources. The pre-earthquake orthophoto was obtained on January 24 2015 (available \nat https://data.linz.govt.nz/layer/52602-canterbury-03m-rural-aerial-photos-2014-2015) and the post-earthquake one on \nDecember 15 2016. The resolutions are 0.3 and 0.2 m, respectively. Table 1: Information about LiDAR data used in this study. 120 \n \nPre-earthquake LiDAR \nPost-earthquake LiDAR \nDate of acquisition \n13/03/2014 – 20/03/2014 \n03/12/2016 – 06/01/2017 \nCommissioned by/provided by \nUSC-UCLA-GNS science/NCALM \nLand Information New Zealand/AAM NZ \nAvailability \nhttps://doi.org/10.5069/G9G44N75 \nOn request \nOriginal point density (pts/m²) \n9.02 \n- \nNumber of ground points \n10,660,089 \n63,729,096 \nGround point density (pts/m²) \n3.8 ± 2.1 \n11.5 ± 6.8 \nVertical accuracy (m) \n0.068 – 0.165 \n0.04 m \nStudy area (m²) \n5,253,133 \n5,253,133 Table 1: Information about LiDAR data used in this study. 120 3. Methods: 3D point cloud differencing to detect and measure landslides 3.1 3D point cloud differencing with M3C2 The method developed here to detect landslides consists in a 3D point cloud differencing between two epochs using the M3C2 \nalgorithm (Lague et al., 2013) available in the Cloudcompare software (Cloudcompare v2.11, 2020). This algorithm estimates \n125 \northogonal distances along the surface normal directly on 3D point clouds without the need for surface interpolation or \ngridding. While M3C2 can be applied on all points, the algorithm can use an accessory point cloud, called core points, of \narbitrary geometry which we impose in our case to be a regular grid with constant horizontal spacing generated by the \nrasterization of one of the two clouds. All the M3C2 calculations will be done in 3D using the raw point clouds, but the results will be “stored” on the core points. The use of a regular grid of core points has four advantages: (i) a regular sampling of the \n130 \nresults which allows to compute robust statistics of changes unbiased by spatial variations in point density; (ii) it facilitates the \nvolume calculation and the uncertainty assessment on total volume; (iii) it can be directly reused with 2D GIS as a raster; and will be “stored” on the core points. The use of a regular grid of core points has four advantages: (i) a regular sampling of the \n130 \nresults which allows to compute robust statistics of changes unbiased by spatial variations in point density; (ii) it facilitates the \nvolume calculation and the uncertainty assessment on total volume; (iii) it can be directly reused with 2D GIS as a raster; and 5 5 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. (iv) it speeds up calculations, although in the proposed workflow, computation time is not an issue and can be done on a regular \nlaptop. (iv) it speeds up calculations, although in the proposed workflow, computation time is not an issue and can be done on a regular \nlaptop. The first step of M3C2 consists in the calculation of a 3D surface normal for each core point at a scale D (called the normal \n135 \nscale) by fitting a plane to the points of the first dataset located within a radius of size D/2 of the core point. 3.1 3D point cloud differencing with M3C2 M3C2 \nhas the option to compute the distance vertically which bypasses the normal calculation, and we use this option several time \nin the workflow. We use the abbreviation vertical-M3C2 in that case and 3D-M3C2 otherwise. where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud a where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud at scale d measured along the where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud at scale d measured along the \nnormal direction, n1 and n2 are the number of points and reg is the co-registration error between the two epochs, assumed \n145 \nspatially uniform and isotropic in our case, but which could be spatially variable and anisotropic (James et al., 2017). M3C2 \nhas the option to compute the distance vertically which bypasses the normal calculation, and we use this option several time \nin the workflow. We use the abbreviation vertical-M3C2 in that case and 3D-M3C2 otherwise. normal direction, n1 and n2 are the number of points and reg is the co-registration error between the two epochs, assumed \n145 \nspatially uniform and isotropic in our case, but which could be spatially variable and anisotropic (James et al., 2017). M3C2 \nhas the option to compute the distance vertically which bypasses the normal calculation, and we use this option several time \nin the workflow. We use the abbreviation vertical-M3C2 in that case and 3D-M3C2 otherwise. 3.2 Parameters selection and 3D point cloud differencing performance In this section, we explain how to select the appropriate normal scale D and projection scale d to detect landslides using M3C2. 150 \nThe normal scale D should be large enough to encompass enough points for a robust calculation, and smooth out small scale \npoint cloud roughness that result in normal orientation flickering and overestimation of the distance between surfaces (Lague \net al., 2013). However, D should be small enough to track the large scale variations in hillslope geometry. By studying \nroughness properties of various natural surface, Lague et al. (2013) proposed a criterion for which the ratio of the normal scale In this section, we explain how to select the appropriate normal scale D and projection scale d to detect landslides using M3C2. 150 \nThe normal scale D should be large enough to encompass enough points for a robust calculation, and smooth out small scale \npoint cloud roughness that result in normal orientation flickering and overestimation of the distance between surfaces (Lague \net al., 2013). However, D should be small enough to track the large scale variations in hillslope geometry. By studying \nroughness properties of various natural surface, Lague et al. (2013) proposed a criterion for which the ratio of the normal scale In this section, we explain how to select the appropriate normal scale D and projection scale d to detect landslides using M3C2. 150 \nThe normal scale D should be large enough to encompass enough points for a robust calculation, and smooth out small scale \npoint cloud roughness that result in normal orientation flickering and overestimation of the distance between surfaces (Lague \net al., 2013). However, D should be small enough to track the large scale variations in hillslope geometry. By studying \nroughness properties of various natural surface, Lague et al. (2013) proposed a criterion for which the ratio of the normal scale and the surface roughness measured at the same scale should be larger than about 25. We thus set D as the minimum scale for \n155 \nwhich a majority of core points verify this criteria (details of the analysis can be found in section S1.1 in the supplements). As \nroughness is a scale and point density dependent measure, we explore a range of normal scales for the pre-earthquake dataset \nwhich has the lowest point density. 3.1 3D point cloud differencing with M3C2 Once the normal \nvectors are defined, the local distance between the two clouds is computed for each core point as the distance of the average \npositions of the two point clouds at a scale d (projection scale). This is done by defining a cylinder of radius d/2, oriented along \nthe normal with a maximum length pmax. Distances are not computed if no intercept is found in the second point cloud. M3C2 135 The first step of M3C2 consists in the calculation of a 3D surface normal for each core point at a scale D (called the normal \n135 \nscale) by fitting a plane to the points of the first dataset located within a radius of size D/2 of the core point. Once the normal \nvectors are defined, the local distance between the two clouds is computed for each core point as the distance of the average \npositions of the two point clouds at a scale d (projection scale). This is done by defining a cylinder of radius d/2, oriented along \nthe normal with a maximum length pmax. Distances are not computed if no intercept is found in the second point cloud. M3C2 also provides uncertainty on the computed distance at 95% of confidence based on local roughness, point density and \n140 \nregistration error as follow: also provides uncertainty on the computed distance at 95% of confidence based on local roughness, point density and \n140 \nregistration error as follow: 𝐿𝑜𝐷95%(𝑑) = ±1.96 (√𝜎1(𝑑)2\n𝑛1\n+ 𝜎2(𝑑)2\n𝑛2\n+ 𝑟𝑒𝑔)\n(2) (2) where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud at scale d measured along the where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud at scale d measured along the \nnormal direction, n1 and n2 are the number of points and reg is the co-registration error between the two epochs, assumed \n5 where LoD95% is the Level of Detection, σ1(d) et σ2(d) are the detrended roughness of each cloud at scale d measured along the \nnormal direction, n1 and n2 are the number of points and reg is the co-registration error between the two epochs, assumed \n5 \nspatially uniform and isotropic in our case, but which could be spatially variable and anisotropic (James et al., 2017). 3.2 Parameters selection and 3D point cloud differencing performance We found that D ~ 10 m represents a threshold scale below which the number of core \npoints verifying the roughness criteria significantly drops. The projection scale d should be chosen such that it is large enough to compute robust statistics using enough points, but small \n160 \nenough to avoid spatial smoothing of the distance measurement. Following Lague et al. (2013), M3C2 computes eq. (2) only \nif 5 points are included in the cylinder of radius d/2 for each cloud. In our case, the pre-EQ data with the lowest point density \nwill in practice set the value of d. We ran 3D-M3C2 with a normal scale D = 10 m and d varying from 1 to 40 meters, and The projection scale d should be chosen such that it is large enough to compute robust statistics using enough points, but small \n160 \nenough to avoid spatial smoothing of the distance measurement. Following Lague et al. (2013), M3C2 computes eq. (2) only \nif 5 points are included in the cylinder of radius d/2 for each cloud. In our case, the pre-EQ data with the lowest point density \nwill in practice set the value of d. We ran 3D-M3C2 with a normal scale D = 10 m and d varying from 1 to 40 meters, and The projection scale d should be chosen such that it is large enough to compute robust statistics using enough points, but small \n160 \nenough to avoid spatial smoothing of the distance measurement. Following Lague et al. (2013), M3C2 computes eq. (2) only \nif 5 points are included in the cylinder of radius d/2 for each cloud. In our case, the pre-EQ data with the lowest point density \nwill in practice set the value of d. We ran 3D-M3C2 with a normal scale D = 10 m and d varying from 1 to 40 meters, and 6 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. found that for d=5 m, eq. (2) can be computed with at least 5 data points for 95 % of the surface. The remaining 5 % of points \ntypically correspond to area of low ground point density below dense vegetation. It was not deemed interesting to increase d \n165 \nabove 5 m for all points as it would deteriorate the ability to detect small landslides. 3.2 Parameters selection and 3D point cloud differencing performance However, to be able to generate M3C2 \nconfidence intervals for as many points as possible, in particular on steep slopes below vegetation, we use a second pass of \nM3C2 with d=10 m using the core points for which no confidence interval was calculated at d=5 m. For this study case, the lower ground point density of the pre-EQ lidar data controls D and d. If the pre-EQ dataset had a point 165 For this study case, the lower ground point density of the pre-EQ lidar data controls D and d. If the pre-EQ dataset had a point \ndensity similar to the post-EQ data, values of D = 4 m and d = 3.5 m could have been used, improving the spatial resolution \n170 \nof the change detection. Finally, the maximum cylinder length pmax was set to 30 m as it allowed to compute the maximum \nchange observed in the study area. This is generally obtained by trial and error. Setting pmax too large, increases significantly \ncomputation time and may result in two different surfaces of the same point cloud being averaged (e.g., near very steep divides \nor in narrow gorges). density similar to the post-EQ data, values of D = 4 m and d = 3.5 m could have been used, improving the spatial resolution \n170 \nof the change detection. Finally, the maximum cylinder length pmax was set to 30 m as it allowed to compute the maximum \nchange observed in the study area. This is generally obtained by trial and error. Setting pmax too large, increases significantly \ncomputation time and may result in two different surfaces of the same point cloud being averaged (e.g., near very steep divides \nor in narrow gorges). 3.3 3D Landslide mapping workflow \n175 Our 3D landslide mapping workflow is divided in four main steps (Fig. 2). Our 3D landslide mapping workflow is divided in four main steps (Fig. 2). Our 3D landslide mapping workflow is divided in four main steps (Fig. 2). Figure 2: Workflow of the landslide detection and volume estimation with schematic representations of the different steps (a,b,c). (a) Schematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b)\nSchematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b) Schematic\n0 \nrepresentation of the segmentation step by connected component. The resulting sources and deposits are individual point clouds\nillustrated in the figure with different colors. (c) Schematic representation of the volume estimation step. Redlines are normal\norientation. (\ng\n) Figure 2: Workflow of the landslide detection and volume estimation with schematic represen Figure 2: Workflow of the landslide detection and volume estimation with schematic representations of the different steps (a,b,c). (a) Schematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b) \nSchematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b) Schematic \nrepresentation of the segmentation step by connected component. The resulting sources and deposits are individual point clouds \nillustrated in the figure with different colors. (c) Schematic representation of the volume estimation step. Redlines are normal \norientation. Figure 2: Workflow of the landslide detection and volume estimation with schematic representations of the different steps (a,b,c). (a) Schematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b) \nSchematic representation of the 3D measurement step with the shadow zone effect. Redlines are normal orientation. (b) Schematic \nrepresentation of the segmentation step by connected component. The resulting sources and deposits are individual point clouds \nillustrated in the figure with different colors. (c) Schematic representation of the volume estimation step. Redlines are normal \norientation. 180 3.3.1. Registration An Iterative Closest Point (ICP) algorithm (Besl and McKay, 1992) is \nperformed on the stable areas, and the obtained rigid transformation is applied to the entire post-earthquake point cloud to \nalign it with the pre-earthquake one. At this stage, the two datasets are considered optimally registered for the stable areas but by landsliding is small. A subsequent 3D-M3C2 calculation is performed to obtain a preliminary map of geomorphic change. 190 \nBy manually selecting a threshold of change, one can identify areas that are assumed to be stable. In our case, this corresponds \nto regions displaying a change smaller than 0.6 m. An Iterative Closest Point (ICP) algorithm (Besl and McKay, 1992) is \nperformed on the stable areas, and the obtained rigid transformation is applied to the entire post-earthquake point cloud to \nalign it with the pre-earthquake one. At this stage, the two datasets are considered optimally registered for the stable areas but 190 with an unknown registration error reg. To estimate this error, the M3C2 algorithm is re-applied on the registered data, and \n195 \nthe registration error is estimated as the standard deviation of the 3D-M3C2 distances measured on stable areas. At this stage, \na 3D map of topographic change is available, but the statistically significant geomorphic change and individual landslides have \nnot been isolated. point clouds, considering only stable areas that are uniformly distributed in the studied region. These were manually selected \n200 \nand represent 23% of the area (Fig. 3). The mean 3D-M3C2 distance is -0.01 m, showing that there is almost no bias in the \nregistration. 95% of the measured distances are within a range of -0.33 to 0.35 m (Fig. 3). The standard deviation is 0.17 m, \nand we thus set reg to 0.17 m. According to eq. (2), LoD95% cannot be smaller than 0.33 m in the ideal case of negligible \nroughness surface. The relatively large registration error in this dataset is related to errors related to flight line imperfect alignment, some classification residual errors (e.g. low vegetation classified as ground) and residual errors related to the ICP \n205 \n(Fig. 3). Figure 3: Map of 3D-M3C2 distances on stable areas used to estimate the registration error and associated histogram. Map is a point \ncloud colored with the pre-earthquake orthophoto (Aerial survey, 2017). Figure 3: Map of 3D-M3C2 distances on stable areas used to estimate the registration error and associated histogram. 3.3.1. Registration Map is a point \ncloud colored with the pre-earthquake orthophoto (Aerial survey, 2017). 3.3.2. Geomorphic change detection \n210 3.3.1. Registration 7\nTo detect geomorphic changes and landslides, the two datasets need to be co-registered as closely as possible and any large-\n185 \nscale tectonic deformation need to be adjusted. No significant intra-dataset deformation was detected, but the produced LiDAR \ndatasets did have a vertical shift larger than 1 m. To correct for it, a grid of core points is first created by rasterizing the dataset To detect geomorphic changes and landslides, the two datasets need to be co-registered as closely as possible and any large-\n185 \nscale tectonic deformation need to be adjusted. No significant intra-dataset deformation was detected, but the produced LiDAR \ndatasets did have a vertical shift larger than 1 m. To correct for it, a grid of core points is first created by rasterizing the dataset 7 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. with the largest point density. Then, a vertical-M3C2 calculation is performed and the mode of the resulting distribution is \nused to adjust the two datasets by a vertical shift of 1.36 m. This approach is valid only when the fraction of the surface affected with the largest point density. Then, a vertical-M3C2 calculation is performed and the mode of the resulting distribution is \nused to adjust the two datasets by a vertical shift of 1.36 m. This approach is valid only when the fraction of the surface affected \nby landsliding is small. A subsequent 3D-M3C2 calculation is performed to obtain a preliminary map of geomorphic change. 190 \nBy manually selecting a threshold of change, one can identify areas that are assumed to be stable. In our case, this corresponds \nto regions displaying a change smaller than 0.6 m. An Iterative Closest Point (ICP) algorithm (Besl and McKay, 1992) is \nperformed on the stable areas, and the obtained rigid transformation is applied to the entire post-earthquake point cloud to \nalign it with the pre-earthquake one. At this stage, the two datasets are considered optimally registered for the stable areas but by landsliding is small. A subsequent 3D-M3C2 calculation is performed to obtain a preliminary map of geomorphic change. 190 \nBy manually selecting a threshold of change, one can identify areas that are assumed to be stable. In our case, this corresponds \nto regions displaying a change smaller than 0.6 m. 3.3.2. Geomorphic change detection \n210 220 3.3.3. Landslide source and deposit segmentation Prior to landslide segmentation, a vertical-M3C2 is performed on the remaining erosion and deposition points located on \nhillslopes in order to estimate landslide volume afterwards. This operation will be described in the next section. Core points \nwith significant change are segmented by a connected component analysis (Fig.2.c). This technique segments a point cloud into compact sub-clouds based on a minimum distance threshold Dm and a minimum number of points per Np sub-clouds \n225 \n(Lumia et al., 1983). A minimum distance of 2 m was chosen as it corresponds to the scale where both the amalgamation effect \nand the over-segmentation are limited (section S2.2.2 in the supplements). Due to the large M3C2 cylinder length pmax required \nto detect deep landslides, small artefacts with a large distance value, near pmax, occur. To ensure that the smallest detected \nlandslides are not affected by these artefacts, the landslide mapping workflow was performed on the subsample versions of the densest point cloud (section 3.2) with a minimum number of points Np set to 1 during the segmentation step. All detected \n230 \nchanges are thus artefacts. The artefact area distribution was then compared with the landslide area distribution using the pre- \nand post-earthquake data (details in section 2.2.1 in the supplements). We finally impose a minimum surface of 20 m², \ncorresponding to 20 points, as it corresponds to the smallest area where artefacts are limited. The final dataset represents all \nthe individualized landslide sources and deposit zones detected. densest point cloud (section 3.2) with a minimum number of points Np set to 1 during the segmentation step. All detected \n230 \nchanges are thus artefacts. The artefact area distribution was then compared with the landslide area distribution using the pre- \nand post-earthquake data (details in section 2.2.1 in the supplements). We finally impose a minimum surface of 20 m², \ncorresponding to 20 points, as it corresponds to the smallest area where artefacts are limited. The final dataset represents all \nthe individualized landslide sources and deposit zones detected. 3.3.4. Landslide volume estimation \n235 \n While 3D normal computation is optimal to detect geomorphic changes, it is not suitable for volume estimation which requires \nto consider normals with parallel directions for a given landslide. As shown in figure 2.a, considering 3D normals can lead to \n“shadow zones”, due to surface roughness, which would result in a biased volume estimate. 3.3.2. Geomorphic change detection \n210 The registration error reg is then used in a first 3D-M3C2, using the pre-determined projection scale d=5m, to estimate the \nspatially variable LoD95% according to eq. (2). Then, a second 3D-M3C2 is performed at a larger projection scale (d=10m) The registration error reg is then used in a first 3D-M3C2, using the pre-determined projection scale d=5m, to estimate the \nspatially variable LoD95% according to eq. (2). Then, a second 3D-M3C2 is performed at a larger projection scale (d=10m) 8 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. only on core points for which a confidence interval could not be estimated due to the insufficient points at d=5m. These points \ngenerally correspond to ground points under canopy on steep slopes, and represented typically 5 % of the core points over the \nentire area, but ~ 15-20 % of steep slopes prone to landsliding. The statistically significant geomorphic changes at the 95% \n215 \nconfidence interval are then obtained by considering core points for which the 3D-M3C2 distance is larger than LoD95%. This \nprocess highlights changes associated to all geomorphic processes, including landsliding, but also erosion and deposition \nprocesses in the fluvial domain. Detected changes located in the river, and specifically related to river dynamics are removed \nmanually to select only landslides sources and deposits. This phase is performed by manual visual inspection to prevent \nlandslide deposited at the bottom of valleys to be removed. 220 only on core points for which a confidence interval could not be estimated due to the insufficient points at d=5m. These points \ngenerally correspond to ground points under canopy on steep slopes, and represented typically 5 % of the core points over the \nentire area, but ~ 15-20 % of steep slopes prone to landsliding. The statistically significant geomorphic changes at the 95% \n215 \nconfidence interval are then obtained by considering core points for which the 3D-M3C2 distance is larger than LoD95%. This \nprocess highlights changes associated to all geomorphic processes, including landsliding, but also erosion and deposition \nprocesses in the fluvial domain. Detected changes located in the river, and specifically related to river dynamics are removed \nmanually to select only landslides sources and deposits. This phase is performed by manual visual inspection to prevent \nl\nd lid d\nit d t th b tt\nf\nll\nt b\nd\n220 215 landslide deposited at the bottom of valleys to be removed. 3.3.3. Landslide source and deposit segmentation Therefore, distances and in turn \nvolumes are computed by using a vertical-M3C2 on a grid of core points corresponding to the significant changes (Fig.2.b). While 3D normal computation is optimal to detect geomorphic changes, it is not suitable for volume estimation which requires \nto consider normals with parallel directions for a given landslide. As shown in figure 2.a, considering 3D normals can lead to \n“shadow zones”, due to surface roughness, which would result in a biased volume estimate. Therefore, distances and in turn \nvolumes are computed by using a vertical-M3C2 on a grid of core points corresponding to the significant changes (Fig.2.b). As the core points are regularly spaced by 1 m, the landslide volume is simply the sum of the vertical-M3C2 distances estimated \n240 \nfrom the individualized landslides. While the distance uncertainty predicted by the vertical-M3C2 could be used as the volume \nuncertainty, it significantly overpredicts the true distance uncertainty due to non-optimal normal orientation for the estimation \nof point cloud roughness on steep slopes (i.e., the roughness is not the detrended roughness). For each landslide source and \ndeposit, we thus compute the volume uncertainty from the sum of the 3D-M3C2 uncertainty measured at each core point, not As the core points are regularly spaced by 1 m, the landslide volume is simply the sum of the vertical-M3C2 distances estimated \n240 \nfrom the individualized landslides. While the distance uncertainty predicted by the vertical-M3C2 could be used as the volume \nuncertainty, it significantly overpredicts the true distance uncertainty due to non-optimal normal orientation for the estimation \nof point cloud roughness on steep slopes (i.e., the roughness is not the detrended roughness). For each landslide source and \ndeposit, we thus compute the volume uncertainty from the sum of the 3D-M3C2 uncertainty measured at each core point, not 9 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. the vertical-M3C2 uncertainty. The volume uncertainty is specific to each landslide sources and deposits and depends on the \n245 \nlocal surface properties such as roughness, the number of point considered and the global registration error, but not on the \nvolume itself. the vertical-M3C2 uncertainty. The volume uncertainty is specific to each landslide sources and deposits and depends on the \n245 \nlocal surface properties such as roughness, the number of point considered and the global registration error, but not on the \nvolume itself. 4.1 Geomorphic change and landslide detection inventory The map of the 3D-M3C2 distances prior to statistically significant change analysis and segmentation highlights erosion (i.e. 250 \nnegative 3D distances) and deposition areas (i.e. positive 3D distances) located on hillslopes and in the river (Fig. 4a). This \nprovide a rare insight into topographic changes following large earthquakes. Most of the detected changes on hillslopes \ncorrespond to mass movements such as landslides and rockfalls with variable sizes, shapes and a high portion of them occurred \non a previous unstable bare-rock zone in the west part of the study area. Small patch of erosion ranging between 10 m² to 100 m² occur over the entire study area. Their large number illustrate how difficult it would be to manually extract them. Most of \n255 \nthe deposit areas are on hillslopes while some the deposits of large landslides have reached the river. The erosion and deposition \nof sediments by fluvial processes can also be observed in the river. The area extent of significant changes, where the absolute amplitude of change is greater than LoD95%, represents 17.5 % of \nthe study area (Fig 4.b). Most points associated to stable areas or artificial changes are correctly filtered by the local confidence interval calculation. In particular, points located under the canopy, or associated to a low point density, or wrongly classified \n260 \nas ground while being vegetation points, leading to locally high values of roughness, result into a higher LoD95% and therefore \nrequires a larger topographic change to be detected as significant. After the manual removal of changes in the fluvial domain \nrelated to fluvial processes, the minimum 3D-M3C2 distance (or minimum LoD95%) on significant change areas is 0.34 ± 0.33 \nm and the maximum is 29.9 ± 0.61 m, both corresponding to erosion areas (Fig.4.b). interval calculation. In particular, points located under the canopy, or associated to a low point density, or wrongly classified \n260 \nas ground while being vegetation points, leading to locally high values of roughness, result into a higher LoD95% and therefore \nrequires a larger topographic change to be detected as significant. After the manual removal of changes in the fluvial domain \nrelated to fluvial processes, the minimum 3D-M3C2 distance (or minimum LoD95%) on significant change areas is 0.34 ± 0.33 \nm and the maximum is 29.9 ± 0.61 m, both corresponding to erosion areas (Fig.4.b). The point cloud of significant changes was segmented to identify the landslide sources (i.e. 4.1 Geomorphic change and landslide detection inventory net erosion) and deposits (i.e. net \n265 \ndeposition). The final landslide inventory, contains a total of 1431 sources and 853 deposits (Fig. 4c). The difference in the \nnumber occurs because many sources share the same deposits that are concentrated at the toe of hillslopes. For sources, the \nmean absolute vertical-M3C2 distance is 2.11 m, the standard deviation 2.62 m and the maximum absolute value 22.8 ± 1.17 \nm. For deposits, the mean absolute vertical-M3C2 distance is 2.37 m, the standard deviation 3.08 m and the maximum absolute \nvalue maximum 29.9 ± 0.61 m. 270 The point cloud of significant changes was segmented to identify the landslide sources (i.e. net erosion) and deposits (i.e. net \n265 \ndeposition). The final landslide inventory, contains a total of 1431 sources and 853 deposits (Fig. 4c). The difference in the \nnumber occurs because many sources share the same deposits that are concentrated at the toe of hillslopes. For sources, the \nmean absolute vertical-M3C2 distance is 2.11 m, the standard deviation 2.62 m and the maximum absolute value 22.8 ± 1.17 \nm. For deposits, the mean absolute vertical-M3C2 distance is 2.37 m, the standard deviation 3.08 m and the maximum absolute \nvalue maximum 29 9 ± 0 61 m\n270 10 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Figure 4: 3D view showing results of three different steps from the general landslide inventory workflow. A) 3D-M3C2 distances \nfrom the geomorphic change detection step. B) Significant changes (>LoD95%). C) Vertical-M3C2 distances of landslide sources \nand deposits inventory with the post-earthquake orthophoto (12-15-2016, Aerial survey, 2017). The number of landslides sources is \n275 \n1431 and deposits is 853. Figure 4: 3D view showing results of three different steps from the general landslide inventory workflow. A) 3D-M3C2 distances \nfrom the geomorphic change detection step. B) Significant changes (>LoD95%). C) Vertical-M3C2 distances of landslide sources \nand deposits inventory with the post-earthquake orthophoto (12-15-2016, Aerial survey, 2017). The number of landslides sources is \n275 \n1431 and deposits is 853. 11 11 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 4.2 Landslide area and volume analysis Recall that the subsequent analysis is based on a grid of core points with 1 m spacing. For each individual landslide, the area \nA was obtained by computing the number of core points inside the sources region. This represents the vertically projected area, \nto be consistent with the existing literature based on 2D studies of landslide statistics. The area of detected landslides ranges \n280 \nfrom 20 to 42,650 m² for sources and from 20 to 33,513 m² for deposits, and the total source and deposit areas are 438,124 and \n376,363 m², respectively. The area distribution of landslide sources is computed as follow (Hovius et al., 1997; Malamud et \nal., 2004): Recall that the subsequent analysis is based on a grid of core points with 1 m spacing. For each individual landslide, the area \nA was obtained by computing the number of core points inside the sources region. This represents the vertically projected area, \nto be consistent with the existing literature based on 2D studies of landslide statistics. The area of detected landslides ranges \n80 \nfrom 20 to 42,650 m² for sources and from 20 to 33,513 m² for deposits, and the total source and deposit areas are 438,124 and Recall that the subsequent analysis is based on a grid of core points with 1 m spacing. For each individual landslide, the area \nA was obtained by computing the number of core points inside the sources region. This represents the vertically projected area, Recall that the subsequent analysis is based on a grid of core points with 1 m spacing. For each i A was obtained by computing the number of core points inside the sources region. This represents to be consistent with the existing literature based on 2D studies of landslide statistics. The area of detected landslides ranges \n280 \nfrom 20 to 42,650 m² for sources and from 20 to 33,513 m² for deposits, and the total source and deposit areas are 438,124 and \n376,363 m², respectively. The area distribution of landslide sources is computed as follow (Hovius et al., 1997; Malamud et \nal., 2004): to be consistent with the existing literature based on 2D studies of landslide statistics. The area of detected landslides ranges \n280 \nfrom 20 to 42,650 m² for sources and from 20 to 33,513 m² for deposits, and the total source and deposit areas are 438,124 and \n376,363 m², respectively. 4.2 Landslide area and volume analysis 1997; Malamud et al., 2004), with an exponent c = - 1.79 ± 0.03 (Fig.5a). We do not observe a rollover at small landslide areas \n290 \nwhich is considered a characteristic of landslide distribution (Guzzetti et al., 2002; Malamud et al., 2004; Malamud and \nTurcotte, 1999). Varying the parameters of the segmentation does not alter this result (section S2.2.2), nor is a rollover visible \nat lower surface area if we reduce the minimum landslide size to 10 m² (section S2.2.1). This behaviour differs from the one \nobserved from the landslide area distribution from Massey et al. (2020) in the Kaikoura region. Landslide volume V was measured with a vertical-M3C2 on landslide areas detected with a 3D-M3C2. The resulting individual \n295 \nlandslide volume ranges from 0.75 ± 9.57 m3 to 171,175 ± 18,593 m3 for source areas, with a total of 908,055 ± 215,640 m3, \nand from 0.75 ± 17.5 m3 to 154,599 ± 15,188 m3 for deposits, with a total of 1,008,626 ± 172,745 m3. The uncertainty on total \nvolume estimation represents 23.7 % for sources and 17.1 % for deposits. The volume distribution of the landslides sources \nwas defined using equation (3) replacing A by the volume V and also exhibit a typical negative power-law scaling (Fig. 5.a) Landslide volume V was measured with a vertical-M3C2 on landslide areas detected with a 3D-M3C2. The resulting individual \n295 \nlandslide volume ranges from 0.75 ± 9.57 m3 to 171,175 ± 18,593 m3 for source areas, with a total of 908,055 ± 215,640 m3, \nand from 0.75 ± 17.5 m3 to 154,599 ± 15,188 m3 for deposits, with a total of 1,008,626 ± 172,745 m3. The uncertainty on total \nvolume estimation represents 23.7 % for sources and 17.1 % for deposits. The volume distribution of the landslides sources \nwas defined using equation (3) replacing A by the volume V and also exhibit a typical negative power-law scaling (Fig. 5.a) Landslide volume V was measured with a vertical-M3C2 on landslide areas detected with a 3D-M3C2. The resulting individual \n295 \nlandslide volume ranges from 0.75 ± 9.57 m3 to 171,175 ± 18,593 m3 for source areas, with a total of 908,055 ± 215,640 m3, \nand from 0.75 ± 17.5 m3 to 154,599 ± 15,188 m3 for deposits, with a total of 1,008,626 ± 172,745 m3. The uncertainty on total \nvolume estimation represents 23.7 % for sources and 17.1 % for deposits. 4.2 Landslide area and volume analysis The area distribution of landslide sources is computed as follow (Hovius et al., 1997; Malamud et \nal., 2004): 𝑝(𝐴) = 1\n𝑁𝐿𝑇\n× 𝛿𝑁𝐿\n𝛿𝐴\n(3) (3) 285 where p(A) is the probability density of a given area within a substantial landslide inventory, NLT is the total number of \nlandslides and A is the landslide source area. δNLT corresponds to the number of landslides with areas between A and A + δA. The landslide area bin widths δA are equal in logarithmic space. where p(A) is the probability density of a given area within a substantial landslide inventory, NLT is the total number of \nlandslides and A is the landslide source area. δNLT corresponds to the number of landslides with areas between A and A + δA. The landslide area bin widths δA are equal in logarithmic space. The area distribution of landslides obeys a power-law scaling relationship consistent with previous studies (e.g., Hovius et al., he area distribution of landslides obeys a power-law scaling relationship consistent with previous st The area distribution of landslides obeys a power-law scaling relationship consistent with previous studies (e.g., Hovius et al., \n1997; Malamud et al., 2004), with an exponent c = - 1.79 ± 0.03 (Fig.5a). We do not observe a rollover at small landslide areas \n290 \nwhich is considered a characteristic of landslide distribution (Guzzetti et al., 2002; Malamud et al., 2004; Malamud and \nTurcotte, 1999). Varying the parameters of the segmentation does not alter this result (section S2.2.2), nor is a rollover visible \nat lower surface area if we reduce the minimum landslide size to 10 m² (section S2.2.1). This behaviour differs from the one \nobserved from the landslide area distribution from Massey et al. (2020) in the Kaikoura region. 1997; Malamud et al., 2004), with an exponent c = - 1.79 ± 0.03 (Fig.5a). We do not observe a rollover at small landslide areas \n290 \nwhich is considered a characteristic of landslide distribution (Guzzetti et al., 2002; Malamud et al., 2004; Malamud and \nTurcotte, 1999). Varying the parameters of the segmentation does not alter this result (section S2.2.2), nor is a rollover visible \nat lower surface area if we reduce the minimum landslide size to 10 m² (section S2.2.1). This behaviour differs from the one \nobserved from the landslide area distribution from Massey et al. (2020) in the Kaikoura region. 4.2 Landslide area and volume analysis The volume distribution of the landslides sources \nwas defined using equation (3) replacing A by the volume V and also exhibit a typical negative power-law scaling (Fig. 5.a) of the form: 𝑝(𝑉) = 𝑑𝑉𝑒. The exponent of the power-law relationship is e = -1.71 ± 0.04. In contrast to landslide source areas \n300 \ndistribution, a roll-over is visible on the landslide volume distribution around 5 to 10 m3. Considering that the minimum \nLoD95% in 3D is 0.34 m, and that the minimum landslide area is 20 m², the minimum volume that we can confidently should \nbe 6.8 m3, a value consistent with the observed rollover. 35 landslides are indeed smaller than 6.8 m3 in our inventory. They \ncorrespond to peculiar cases of very small landslides in which negative 3D distances close to the LoD95% are positive when measured vertically and thus reduce the apparent volume of eroded material. Volume estimation in these cases should be done \n305 \nlocally in 3D, but requires further development not deemed necessary given that it is restricted to the smallest and shallowest \nlandslides of the inventory. With a direct measurement of landslide volume, it is possible to compute the volume-area relationship (eq.(1); Simonett, 1967; \nLarsen et al., 2010) and to compare it with pre-existing results in New Zealand (Larsen et al., 2010, Massey et al., 2020). Here measured vertically and thus reduce the apparent volume of eroded material. Volume estimation in these cases should be done \n305 \nlocally in 3D, but requires further development not deemed necessary given that it is restricted to the smallest and shallowest \nlandslides of the inventory. With a direct measurement of landslide volume, it is possible to compute the volume-area relationship (eq.(1); Simonett, 1967; \nLarsen et al., 2010) and to compare it with pre-existing results in New Zealand (Larsen et al., 2010, Massey et al., 2020). Here With a direct measurement of landslide volume, it is possible to compute the volume-area relationship (eq.(1); Simonett, 1967; \nLarsen et al., 2010) and to compare it with pre-existing results in New Zealand (Larsen et al., 2010, Massey et al., 2020). Here 12 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. we determine V-A scaling coefficients using two methods: by fitting a linear model (1) on log-transformed data and (2) on \n310 \nlog-binned data. 4.2 Landslide area and volume analysis (2010) relationships derived from soil landslides and from mixed \nsoil landslides and bedrock landslides, respectively (Table 2). However, R² of 0.61 and 0.72 are obtained when considering \nthe parameters of the V-A relationships, derived by Massey et al. (2020), based on the 8,442 cleaned landslides or only soil-\ndominated ones, respectively, of the Kaikoura region. At first order, the V-A relationships we obtained are thus consistent with previous studies. Yet, the V-A scaling relationship obtained with log-binned data best predicts the volume directly measured \n320 \nfrom difference of LiDAR point clouds (Table 2). If the relationships from Larsen et al. (2010) and Massey et al. (2020) were \napplied to our landslide area inventory, the total volume would vary from 0.541x106 m3 to 1.347x106 m3 (Table 2), compared \nto 0.908x106 ± 0.215x106 m3 that we estimate directly. This highlights the overarching sensitivity of the total volume of eroded \nmaterial to the V-A relationship (Li et al., 2014; Marc and Hovius, 2015). The closest evaluation of the total volume is based previous studies. Yet, the V-A scaling relationship obtained with log-binned data best predicts the volume directly measured \n320 \nfrom difference of LiDAR point clouds (Table 2). If the relationships from Larsen et al. (2010) and Massey et al. (2020) were \napplied to our landslide area inventory, the total volume would vary from 0.541x106 m3 to 1.347x106 m3 (Table 2), compared \nto 0.908x106 ± 0.215x106 m3 that we estimate directly. This highlights the overarching sensitivity of the total volume of eroded \nmaterial to the V-A relationship (Li et al., 2014; Marc and Hovius, 2015). The closest evaluation of the total volume is based on the Massey et al. (2020) V-A relationship for soil landslides, that predicts a total volume of 0.934x106 m3. However, this \n325 \nV-A relationship gets close to the total landslide volume by significantly overpredicting the volume of small landslides. The \nopposite is true for the Larsen et al. (2010) V-A relationship for all landslides (1.347x106 m3) which overpredicts the volume \nof large landslides, while their soil landslides relationships only predict half of the total volume. This latter results hints at the \ntransition from shallow dominated landsliding at small areas to deeper bedrock landsliding that may obey a different V-A on the Massey et al. (2020) V-A relationship for soil landslides, that predicts a total volume of 0.934x106 m3. 4.2 Landslide area and volume analysis While the first method leads to a V-A relationship best describing the volume of each landslide, the second \none is not affected by the varying number of landslides in each range of landslide area and leads to a V-A relationship that best \nmatches the total landslide volume. Using the first approach, we find a volume-area scaling exponent of 𝛾= 1.18 ± 0.01 and \nan intercept log 𝛼= −0.42 ± 0.02 m64 with a determination coefficient 𝑅2 = 0.99 (Fig.5.c). Using the second method, we we determine V-A scaling coefficients using two methods: by fitting a linear model (1) on log-transformed data and (2) on \n310 \nlog-binned data. While the first method leads to a V-A relationship best describing the volume of each landslide, the second \none is not affected by the varying number of landslides in each range of landslide area and leads to a V-A relationship that best \nmatches the total landslide volume. Using the first approach, we find a volume-area scaling exponent of 𝛾= 1.18 ± 0.01 and \nan intercept log 𝛼= −0.42 ± 0.02 m64 with a determination coefficient 𝑅2 = 0.99 (Fig.5.c). Using the second method, we \nfind 𝛾= 1.20 ± 0.02, an intercept log 𝛼= −0.41 ± 0.07 m0.6 and a determination coefficient 𝑅2 = 0.99. We also obtain a \n315 \ngood correlation R² of 0.86 and 0.82 with the Larsen et al. (2010) relationships derived from soil landslides and from mixed \nsoil landslides and bedrock landslides, respectively (Table 2). However, R² of 0.61 and 0.72 are obtained when considering \nthe parameters of the V-A relationships, derived by Massey et al. (2020), based on the 8,442 cleaned landslides or only soil-\ndominated ones, respectively, of the Kaikoura region. At first order, the V-A relationships we obtained are thus consistent with \nprevious studies. Yet, the V-A scaling relationship obtained with log-binned data best predicts the volume directly measured \n320 \nfrom difference of LiDAR point clouds (Table 2). If the relationships from Larsen et al. (2010) and Massey et al. (2020) were \napplied to our landslide area inventory, the total volume would vary from 0.541x106 m3 to 1.347x106 m3 (Table 2), compared \nto 0.908x106 ± 0.215x106 m3 that we estimate directly. This highlights the overarching sensitivity of the total volume of eroded \nmaterial to the V-A relationship (Li et al., 2014; Marc and Hovius, 2015). 4.2 Landslide area and volume analysis The closest evaluation of the total volume is based \non the Massey et al. (2020) V-A relationship for soil landslides, that predicts a total volume of 0.934x106 m3. However, this \n325 \nV-A relationship gets close to the total landslide volume by significantly overpredicting the volume of small landslides. The \nopposite is true for the Larsen et al. (2010) V-A relationship for all landslides (1.347x106 m3) which overpredicts the volume \nof large landslides, while their soil landslides relationships only predict half of the total volume. This latter results hints at the \ntransition from shallow dominated landsliding at small areas to deeper bedrock landsliding that may obey a different V-A \nrelationship at large area. Our inventory is not large enough to confidently define a specific trend for deep and large landslides. 330 an intercept log 𝛼\n0.42 ± 0.02 m with a determination coefficient 𝑅\n0.99 (Fig.5.c). Using the second method, we \nfind 𝛾= 1.20 ± 0.02, an intercept log 𝛼= −0.41 ± 0.07 m0.6 and a determination coefficient 𝑅2 = 0.99. We also obtain a \n315 \ngood correlation R² of 0.86 and 0.82 with the Larsen et al. (2010) relationships derived from soil landslides and from mixed \nsoil landslides and bedrock landslides, respectively (Table 2). However, R² of 0.61 and 0.72 are obtained when considering \nthe parameters of the V-A relationships, derived by Massey et al. (2020), based on the 8,442 cleaned landslides or only soil-\ndominated ones, respectively, of the Kaikoura region. At first order, the V-A relationships we obtained are thus consistent with find 𝛾= 1.20 ± 0.02, an intercept log 𝛼= −0.41 ± 0.07 m0.6 and a determination coefficient 𝑅2 = 0.99. We also obtain a \n315 \ngood correlation R² of 0.86 and 0.82 with the Larsen et al. (2010) relationships derived from soil landslides and from mixed \nsoil landslides and bedrock landslides, respectively (Table 2). However, R² of 0.61 and 0.72 are obtained when considering \nthe parameters of the V-A relationships, derived by Massey et al. (2020), based on the 8,442 cleaned landslides or only soil-\ndominated ones, respectively, of the Kaikoura region. At first order, the V-A relationships we obtained are thus consistent with find 𝛾= 1.20 ± 0.02, an intercept log 𝛼= −0.41 ± 0.07 m0.6 and a determination coefficient 𝑅2 = 0.99. We also obtain a \n315 \ngood correlation R² of 0.86 and 0.82 with the Larsen et al. 4.2 Landslide area and volume analysis However, this \n325 \nV-A relationship gets close to the total landslide volume by significantly overpredicting the volume of small landslides. The \nopposite is true for the Larsen et al. (2010) V-A relationship for all landslides (1.347x106 m3) which overpredicts the volume \nof large landslides, while their soil landslides relationships only predict half of the total volume. This latter results hints at the \ntransition from shallow dominated landsliding at small areas to deeper bedrock landsliding that may obey a different V-A relationship at large area. Our inventory is not large enough to confidently define a specific trend for deep and large landslides. 330 13 13 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Figure 5: Landslide sources inventory analysis of the study area (NLT=1431). Landslide area (a) and volume (b) probability density \ndistribution. (a) also contains the original V2 landslide inventory area distribution of Massey et al. (2020). Power-law fits begin \nrespectively at 20 m² and 20 m3. (b) Landslide probability density in function of landslide volume. (c) Volume-area power-law scaling \nrelationship with uncertainty on volume and comparison with Larsen et al. (2010) and Massey et al. (2020) relationships obtained \nin New Zealand. All scaling parameter values are summarized in Table 2. Grey lines are depth vs area relationship for different \nmean depths. Figure 5: Landslide sources inventory analysis of the study area (NLT=1431). Landslide area (a) and volume (b) probability density \ndistribution. (a) also contains the original V2 landslide inventory area distribution of Massey et al. (2020). Power-law fits begin \nrespectively at 20 m² and 20 m3. (b) Landslide probability density in function of landslide volume. (c) Volume-area power-law scaling \nrelationship with uncertainty on volume and comparison with Larsen et al. (2010) and Massey et al. (2020) relationships obtained \nin New Zealand. All scaling parameter values are summarized in Table 2. Grey lines are depth vs area relationship for different \nmean depths. Table 2: Power-law scaling parameter values of the relations show in figure 5. Log α and γ are scaling parameters from the landslide \narea-volume relationship. Unit of α is [L(3-2γ)] with L in meters. Landslide source area and volume distribution coefficients are b \n340 \nand d while exponents are c and e respectively. The coefficient of determination R² is also given for each power-law fit function. 4.2 Landslide area and volume analysis The \ntotal volume corresponds to the application of a specific V-A relationship to the landslide areas of our inventory. log b-d / log α \nc-e / γ \nR² \nTotal Volume m3 \nLandslide area distribution \n0.75 ± 0.09 \n-1.79 ± 0.03 \n0.99 \n- \nLandslide volume distribution \n0.41 ± 0.14 \n-1.71 ± 0.04 \n0.99 \n0.908 × 106 \n(direct measurement) \nV-A relationship from averaged data (this study) \n-0.41 ± 0.07 \n1.20 ± 0.02 \n0.99 \n0.900 × 106 \nV-A relationship from raw data (this study) \n-0.42 ± 0.02 \n1.18 ± 0.01 \n0.86 \n0.740 × 106 \nV-A relationship for soil landslides \n(Larsen et al., 2010) \n-0.37 ± 0.06 \n1.13 ± 0.03 \n0.86 \n0.541 × 106 \nV-A relationship for mixed soil and bedrock landslide \n(Larsen et al., 2010) \n-0.86 ± 0.05 \n1.36 ± 0.01 \n0.82 \n1.347 × 106 Table 2: Power-law scaling parameter values of the relations show in figure 5. Log α and γ are scaling parameters from the landslide \narea-volume relationship. Unit of α is [L(3-2γ)] with L in meters. Landslide source area and volume distribution coefficients are b \n340 \nand d while exponents are c and e respectively. The coefficient of determination R² is also given for each power-law fit function. The \ntotal volume corresponds to the application of a specific V-A relationship to the landslide areas of our inventory. 14 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. V-A relationship for soil landslides \n (Massey et al., 2020) \n0.12 ± 0.04 \n1.06 ± 0.02 \n0.61 \n0.934 × 106 \nV-A relationship for all landslides \n (Massey et al., 2020) \n-0.05 ± 0.02 \n1.109 ± 0.008 \n0.72 \n0.947 × 106 V-A relationship for soil landslides \n (Massey et al., 2020) \nV-A relationship for all landslides \n (Massey et al., 2020) 4.3 Reactivated landslides and new failures Because the 3D measurement approach only depends on local topographic change, we evaluate the fraction of reactivated Because the 3D measurement approach only depends on local topographic change, we evaluate the fraction of reactivated \nlandslides in the population that would have been hard or impossible to detect with 2D imagery based on texture change \n45 Because the 3D measurement approach only depends on local topographic change, we eval Because the 3D measurement approach only depends on local topographic change, we evaluate the fraction of reactivated \nlandslides in the population that would have been hard or impossible to detect with 2D imagery based on texture change \n345 \n(Fig.6.a). We hereinafter distinguish between new and reactivated landslides, by considering that reactivated landslides occur \non bare rock areas in the pre-earthquake imagery or on vegetated areas with limited texture and colour contrast between the \ntwo epochs. These definitions were chosen following classical approaches for landslide detection based on vegetation cover \nanalysis (Behling et al., 2014; Marc et al., 2019; Martha et al., 2010; Massey et al., 2018; Pradhan et al., 2016). landslides in the population that would have been hard or impossible to detect with 2D imagery based on texture change \n345 \n(Fig.6.a). We hereinafter distinguish between new and reactivated landslides, by considering that reactivated landslides occur \non bare rock areas in the pre-earthquake imagery or on vegetated areas with limited texture and colour contrast between the \ntwo epochs. These definitions were chosen following classical approaches for landslide detection based on vegetation cover \nanalysis (Behling et al., 2014; Marc et al., 2019; Martha et al., 2010; Massey et al., 2018; Pradhan et al., 2016). Most of the detected landslides are new failures with a total area of 318,726 m² and volume of 636,359 ± 163,496 m3 (Fig.6.a). 350 \nWe find that reactivated landslides have a total area 119,398 m² and volume of 271,695 ± 52,144 m3, which represents 27.2 % \nand 29.9 ± 12.8 % of the total landslide area and volume, respectively (Table 3). Figure 6b illustrates in detail an area \nexperiencing active rock avalanches and debris flow in the pre-earthquake imagery, for which the change in contrast and \nshadows is likely too complex to detect a topographic change from a texture based analysis. On the contrary, the 3D change detection shows that landslide erosion is pervasive in this sector, and corresponds indeed, to the largest landslide detected by \n355 \nour approach. 4.3 Reactivated landslides and new failures The centroids of the landslide inventory (version 1.0) mapped by Massey et al. (2018) \nare shown in comparison (n=27). B) Zoom on a reactivation zone, from the left to the right: pre-earthquake orthophoto (January \n24, 2015), post-earthquake orthophoto (December 15, 2016) and detected reactivation area (only source area) super-imposed to the \npre-earthquake orthophoto (Aerial survey, 2017). 5 365 Table 3: Area and associated volume of the considered new failure and reactivation zones. New failures \n% Total \nReactivation \n% Total \nArea (m²) \n318,726 \n69.9 \n119,398 \n27.2 \nVolume (m3) \n636,359 ± 163,496 \n70.1 ± 34.6 \n271,695 ± 52,144 \n29.9 ± 12.8 Table 3: Area and associated volume of the considered new failure and reactivation zones. 4.3 Reactivated landslides and new failures The difficulty to detect small and reactivated landslides is illustrated by plotting the version 1.0 of the landslide \ninventory from Massey et al. (2018) in our study area (Fig.6.a), which is a database manually validated, and constantly updated. Only 27 landslide sources were initially mapped and none in reactivated zones, where we found 1431 landslides for which \n27.2% of the total area are reactivated. detection shows that landslide erosion is pervasive in this sector, and corresponds indeed, to the largest landslide detected by \n355 \nour approach. The difficulty to detect small and reactivated landslides is illustrated by plotting the version 1.0 of the landslide \ninventory from Massey et al. (2018) in our study area (Fig.6.a), which is a database manually validated, and constantly updated. Only 27 landslide sources were initially mapped and none in reactivated zones, where we found 1431 landslides for which \n27.2% of the total area are reactivated. 15 Figure 6: New failure and reactivation areas identified from detected landslide sources. A) Map of landslide source areas colorized \naccording to new failure or reactivation zone. The centroids of the landslide inventory (version 1.0) mapped by Massey et al. (2018) \nare shown in comparison (n=27). B) Zoom on a reactivation zone, from the left to the right: pre-earthquake orthophoto (January \n24, 2015), post-earthquake orthophoto (December 15, 2016) and detected reactivation area (only source area) super-imposed to the \npre-earthquake orthophoto (Aerial survey, 2017). Table 3: Area and associated volume of the considered new failure and reactivation zones. New failures \n% Total \nReactivation \n% Total \nhttps://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Preprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 360 Figure 6: New failure and reactivation areas identified from detected landslide sources. A) Map of landslide source areas colorized \naccording to new failure or reactivation zone. The centroids of the landslide inventory (version 1.0) mapped by Massey et al. (2018) \nare shown in comparison (n=27). B) Zoom on a reactivation zone, from the left to the right: pre-earthquake orthophoto (January \n24, 2015), post-earthquake orthophoto (December 15, 2016) and detected reactivation area (only source area) super-imposed to the \npre-earthquake orthophoto (Aerial survey, 2017). 65 Figure 6: New failure and reactivation areas identified from detected landslide sources. A) Map of landslide source areas colorized \naccording to new failure or reactivation zone. 5.1 3D point cloud differencing and landslide detection \n375 3D point cloud differencing methods have already been applied in previous studies to detect geomorphic changes on single \nlandslides using point clouds obtained by photogrammetry with drone-based images (Esposito et al., 2017; Stumpf et al., 2014, \n2015). Larger-scale approaches have been attempted to evaluate the impact of a given landslide (Bossi et al., 2015). However, \nthe DoD, based on gridded data, remains the dominant approach to evaluate topographic changes in the context of glacier \ndynamics, fluvial dynamics or tectonic deformation analysis (e.g. Passalacqua et al., 2015 for a review). To our knowledge, \n380 \nthe systematic detection and segmentation of hundreds of landslides from 3D point cloud have not yet been attempted. 5. Discussion The aim of this paper is to investigate the potential of methods based on 3D point cloud differencing to provide a landslide \ninventory map at a region scale from LiDAR data, a total landslide volume estimate and to overcome issues such as landslide \n370 \namalgamation effects on total estimated landslide volume, under-detection of reactivated landslides in 2D imagery analysis as \nwell as limitations of the DoD approach on steep slopes. Here we first discuss the 3D workflow we have developed in 16 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. comparison to traditional DoD approach, and then discuss the benefits of 3D change detection for landslide inventories creation \ncompared to 2D imagery landslide detection. comparison to traditional DoD approach, and then discuss the benefits of 3D change detection for landslide inventories creation \ncompared to 2D imagery landslide detection. 5.1.1 Vertical versus 3D change detection capability, and the M3C2 algorithm However, the 3D approach results in a standard deviation, σ=0.05 m, four times smaller than using a vertical differencing, σ= \n390 \n0.20 m. The map of distance shows that vertical differencing systematically results in much higher distances on steep slopes \nthan the 3D approach, while they both yield similar, low distances, on horizontal surfaces. We thus find that 3D point cloud differencing offers a greater sensitivity to detect changes compared to classical vertical DoD. This difference is particularly important as it propagates into a lower level of detection and uncertainty on volume calculation. Using the M3C2 algorithm in 3D (Lague et al., 2013) also offers the benefit of accounting for spatially variable point density \n395 \nand roughness in estimating a distance uncertainty for each core point, that can subsequently be used in volume uncertainty \ncalculation. For instance, this approach leads to a reduced detection sensitivity in vegetated areas due to a lower ground point \ndensity and potentially higher roughness due to vegetation misclassification. By using a regular grid of core points as in Wagner \net al. (2017), our workflow combines the benefits of working directly with the raw unorganized 3D data, as opposed to DoD Using the M3C2 algorithm in 3D (Lague et al., 2013) also offers the benefit of accounting for spatially variable point density \n395 \nand roughness in estimating a distance uncertainty for each core point, that can subsequently be used in volume uncertainty \ncalculation. For instance, this approach leads to a reduced detection sensitivity in vegetated areas due to a lower ground point \ndensity and potentially higher roughness due to vegetation misclassification. By using a regular grid of core points as in Wagner \net al. (2017), our workflow combines the benefits of working directly with the raw unorganized 3D data, as opposed to DoD where the relationship with the underlying higher point density data is lost, while producing a result with regular sampling that \n400 \ncan easily be used for unbiased spatial statistics, volume calculation and easy integration into 2D GIS software. Compared to \nDoD, if an interpolation is needed, it is performed on the results rather than on the original DEM which can lead to uncontrolled \nerror budget management. 5.1.1 Vertical versus 3D change detection capability, and the M3C2 algorithm The importance of detecting change in 3D, as opposed to vertically in steep slopes can be illustrated by a simple exercise, \nsimilar to section 3.2, in which we use two random sub-sampled versions of the post-event point cloud that represent exactly The importance of detecting change in 3D, as opposed to vertically in steep slopes can be illustrated by a simple exercise, \nsimilar to section 3.2, in which we use two random sub-sampled versions of the post-event point cloud that represent exactly \nthe same surface without any registration error, have similar point density but different sampling of the surface. We apply a \n385 \nvertical-M3C2 and a 3D-M3C2 to the two point clouds, and the maps of change and distance distributions are shown on figure \n7a. Typical of change measurement methods on rough surfaces with random point sampling (e.g., Lague et al., 2013), a non-\nnull distance is often measured even though the two point clouds are samples of exactly the same surface. The distribution of \nmeasured distances is centred near zero, with a mean of −2 .10−4 and 10−4 m, for the vertical and 3D approach respectively. the same surface without any registration error, have similar point density but different sampling of the surface. We apply a \n385 \nvertical-M3C2 and a 3D-M3C2 to the two point clouds, and the maps of change and distance distributions are shown on figure \n7a. Typical of change measurement methods on rough surfaces with random point sampling (e.g., Lague et al., 2013), a non-\nnull distance is often measured even though the two point clouds are samples of exactly the same surface. The distribution of \nmeasured distances is centred near zero, with a mean of −2 .10−4 and 10−4 m, for the vertical and 3D approach respectively. However, the 3D approach results in a standard deviation, σ=0.05 m, four times smaller than using a vertical differencing, σ= \n390 \n0.20 m. The map of distance shows that vertical differencing systematically results in much higher distances on steep slopes \nthan the 3D approach, while they both yield similar, low distances, on horizontal surfaces. We thus find that 3D point cloud differencing offers a greater sensitivity to detect changes compared to classical vertical DoD. This difference is particularly important as it propagates into a lower level of detection and uncertainty on volume calculation. 5.1.2 Tectonic internal deformation, data quality and point clouds registration 5.1.2 Tectonic internal deformation, data quality and point clouds registration One of the most critical part of any 3D point cloud processing method is the co-registration o p\ny\np\np\ng\ng\np\n,\np\nin the context of co-seismic landsliding. With LiDAR data, the registration error will generally set the minimum detectable \n410 \nchange on bare planar surfaces. Here, a rigid translation has been applied on the entire area using an ICP algorithm (Besl and \nMcKay, 1992), de facto assuming that internal deformation during the earthquake was negligible. After applying a vertical \ndisplacement of 1.36 m, we did not observe a systematic horizontal shift of the difference map either north or south of the \nHope fault. We thus conclude that the internal deformation was below the typical registration error in our study area. For larger studied regions with internal deformation and in the absence of a 3D co-seismic deformation model that could be applied to \n415 \nthe post-EQ point cloud (e.g., Massey et al., 2020), our workflow should be applied in a piecewise manner with boundaries \ncorresponding to the main identified faults or deformation zones. For landslide inventories following climatic events, the \napplication to very large dataset should be straightforward as no internal deformation is expected. Similarly, we also noted an \ninternal flight line height mismatch of 0.05-0.12 m in the pre-EQ survey, that was difficult to correct after data delivery and generated some apparent large scale low amplitude topographic change (Fig. 3). This highlights the need for detailed quality \n420 \ncontrol (e.g., by applying M3C2 on overlapping lines) to ensure the highest accuracy possible of the LiDAR data. 5.1.1 Vertical versus 3D change detection capability, and the M3C2 algorithm where the relationship with the underlying higher point density data is lost, while producing a result with regular sampling that \n400 \ncan easily be used for unbiased spatial statistics, volume calculation and easy integration into 2D GIS software. Compared to \nDoD, if an interpolation is needed, it is performed on the results rather than on the original DEM which can lead to uncontrolled \nerror budget management. 17 Figure 7: Comparison between 2D vertical differencing (vertical-M3C2) and 3D differencing (3D-M3C2) on the post-EQ sub-\n5 \nsampled randomly two times to generate two point clouds of the same surface with a different sampling. A) Resulting change \ndetection maps of the two different techniques. B) Histogram of the computed distances with the two techniques. https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Figure 7: Comparison between 2D vertical differencing (vertical-M3C2) and 3D differencing (3D-M3C2) on the post-EQ sub-\n405 \nsampled randomly two times to generate two point clouds of the same surface with a different sampling. A) Resulting change \ndetection maps of the two different techniques. B) Histogram of the computed distances with the two techniques. 5.1.4 Landslide topographic changes and feature tracking \n435 Our approach focuses on landslide detection and volume calculation. The workflow we have designed is not suitable for \ndeformation measurement based on feature tracking (Passalacqua et al., 2015). Except for a few landslides with limited \ndisplacement in which point cloud features could be potentially tracked, the severity of landsliding and the long runout of \nmany landslides preclude any attempt in tracking features. Our approach may miss translational landslides on planar hillslopes Our approach focuses on landslide detection and volume calculation. The workflow we have designed is not suitable for \ndeformation measurement based on feature tracking (Passalacqua et al., 2015). Except for a few landslides with limited \ndisplacement in which point cloud features could be potentially tracked, the severity of landsliding and the long runout of \nmany landslides preclude any attempt in tracking features. Our approach may miss translational landslides on planar hillslopes \nfor which topographic change occurs in the direction of the surface normal. Given that hillslopes are generally not perfectly \n440 \nplanar, any significant translation parallel to the hillslope will generate a topographic change, especially in the source area. As \nh\ni\nfi\n8b\nh d\nl i\nl l\nd lid\nl h\nh i\nh d f\ni\nll l\nh for which topographic change occurs in the direction of the surface normal. Given that hillslopes are generally not perfectly \n440 \nplanar, any significant translation parallel to the hillslope will generate a topographic change, especially in the source area. As \nshown in fig. 8b, our approach detects translational landslides, although it cannot compute the deformation parallel to the \nhillslopes. The only element that could be easily tracked in this inventory are the barycenter of the source and associated \ndeposit of each landslide, to explore runout dynamics, but we have not investigated this option yet. for which topographic change occurs in the direction of the surface normal. Given that hillslopes are generally not perfectly \n440 \nplanar, any significant translation parallel to the hillslope will generate a topographic change, especially in the source area. As \nshown in fig. 8b, our approach detects translational landslides, although it cannot compute the deformation parallel to the \nhillslopes. The only element that could be easily tracked in this inventory are the barycenter of the source and associated \ndeposit of each landslide, to explore runout dynamics, but we have not investigated this option yet. 5.1.3 Landslide segmentation Another important aspect of the method is the segmentation procedure to individualize sources and deposits. We performed a \nsensitivity analysis to evaluate the impact of the minimum distance between sub-clouds Dm, on the number of segmented \nlandslides and related geometric characteristics (section 3.2, Suppl. Material S.2.2.2). We show that below 4 m, the choice of \n425 Another important aspect of the method is the segmentation procedure to individualize sources and deposits. We performed a \nsensitivity analysis to evaluate the impact of the minimum distance between sub-clouds Dm, on the number of segmented landslides and related geometric characteristics (section 3.2, Suppl. Material S.2.2.2). We show that below 4 m, the choice of \n425 18 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Dm has little impact on the pdf of area, pdf of volume, and the volume-area relationship. However, above 4 m, amalgamation \nstarts to significantly alter the landslide geometry statistics. A similar analysis should be performed for each new dataset to \nevaluate the best segmentation scale. The connected component segmentation is a simple and rapid way to individualize \nlandslides but given the complexity of the 3D dataset, and in particular the very large range of landslide sizes, inevitably \nexhibits some drawbacks and is subject to improvement. For instance, landslides occurring on both side of a same collapsed \n430 \ndivide are considered as one landslide if they are close enough (Fig. 8.a). More advanced segmentation approaches accounting \nfor normal direction, divide organization and 3D depth maps of amalgamated scars would be needed to improve the \nsegmentation and get more robust results on very large datasets. We note however that these issues do not affect the total \nlandslide volume calculation. Dm has little impact on the pdf of area, pdf of volume, and the volume-area relationship. However, above 4 m, amalgamation \nstarts to significantly alter the landslide geometry statistics. A similar analysis should be performed for each new dataset to \nevaluate the best segmentation scale. The connected component segmentation is a simple and rapid way to individualize \nlandslides but given the complexity of the 3D dataset, and in particular the very large range of landslide sizes, inevitably exhibits some drawbacks and is subject to improvement. For instance, landslides occurring on both side of a same collapsed \n430 \ndivide are considered as one landslide if they are close enough (Fig. 8.a). 5.1.3 Landslide segmentation More advanced segmentation approaches accounting \nfor normal direction, divide organization and 3D depth maps of amalgamated scars would be needed to improve the \nsegmentation and get more robust results on very large datasets. We note however that these issues do not affect the total \nlandslide volume calculation. 5.2.1 Volume of landslide sources and deposits \n450 Over the studied area of ~5 km2, 1431 landslide sources and 853 landslide deposits were detected with the 3D point cloud \nprocessing workflow. This relatively large number of landslides is mostly associated to the large number of small landslides \n(< 100 m², n = 977) that were detected thanks to the resolution of the data. The scaling of the pdf of volume of sources, e = -\n1.71, indicates a slight tendency for the overall eroded volume to be dominated by the largest landslide (171,175 m3, that is 18.8 % of the total volume). The uncertainty on total landslide volume, 17.1% to 23.7 % for deposits and sources, respectively, \n455 \nmight appear large, but is based on a conservative 95% confidence interval that we use throughout our analysis. It is mostly \ndominated by the registration error (reg = 0.17 m) and by the lower point cloud density of the pre-earthquake LiDAR data \n(Table.1). Within this uncertainty, the total volume of sources (908,055 ± 215,640 m3) and deposits (1,008,626 ± 172,745 m3) \nare not statistically different. The larger volume of deposit is however consistent with rock decompaction during landsliding. We also note that debris deposits form more concentrated and thicker patches at the toe of hillslopes, and are thus more \n460 \nsystematically above the detection threshold. Very shallow rockfalls might not be detected and accounted for in the source \nvolume. Hence, we expect to detect more of the population of landslide deposits than of the population of sources. We also note that debris deposits form more concentrated and thicker patches at the toe of hillslopes, and are thus more \n460 \nsystematically above the detection threshold. Very shallow rockfalls might not be detected and accounted for in the source \nvolume. Hence, we expect to detect more of the population of landslide deposits than of the population of sources. 5.1.4 Landslide topographic changes and feature tracking \n435 Figure 8: Two different point of interest of the landslide inventory results. a) Zoom to the biggest landslide of the inventory showing \namalgamation across the divide. b) Detected translational sliding of a part of the hillslope. The point cloud is superimposed with \nthe post-earthquake orthophoto (December 15, 2016; Aerial survey, 2017) 445 Figure 8: Two different point of interest of the landslide inventory results. a) Zoom to the biggest landslide of the inventory showing \namalgamation across the divide. b) Detected translational sliding of a part of the hillslope. The point cloud is superimposed with \nthe post-earthquake orthophoto (December 15, 2016; Aerial survey, 2017) 19 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 5.2.2 Distribution of landslide volume and area: power-law behaviour We obtain a range of landslide area over 3 to 4 orders of magnitude (20 to 42,650 m ) for which we constructed the pdf of area \nand volume. In landslide analysis, the pdf of landslide area represents the basis to estimate large-scale landslide erosion (Larsen \n465 \net al. 2010). This distribution has generally a negative power law behaviour for landslide areas larger than a given threshold \nand displays a rollover for smaller landslides (Fan et al. 2019; Guzzetti et al., 2002; Malamud et al., 2004; Malamud and \nTurcotte, 1999; Stark and Hovius, 2001). In this study, we find that the exponent c of the power-law for the landslide area \ndistribution is -1.79 ± 0.03 (Fig.5a). This is roughly consistent with the exponents obtained over the entire Kaikoura coseismic and volume. In landslide analysis, the pdf of landslide area represents the basis to estimate large-scale landslide erosion (Larsen \n465 \net al. 2010). This distribution has generally a negative power law behaviour for landslide areas larger than a given threshold \nand displays a rollover for smaller landslides (Fan et al. 2019; Guzzetti et al., 2002; Malamud et al., 2004; Malamud and \nTurcotte, 1999; Stark and Hovius, 2001). In this study, we find that the exponent c of the power-law for the landslide area \ndistribution is -1.79 ± 0.03 (Fig.5a). This is roughly consistent with the exponents obtained over the entire Kaikoura coseismic landslide inventory of -1.88 (NLT = 10,195; Massey et al., 2018) and more recently of -2.10 (NLT = 29,557;Massey et al., 2020). 470 \nWe present here, one of the first landslide volume distribution derived directly from 3D topographic data, rather than inferred \nfrom the combination of detection of the landslide area distribution on 2D data and an estimated volume-area relationship \nwhich is much more difficult to precisely estimate. Our direct measurements show that the landslide volume distribution indeed \nobeys a power-law relationship with an exponent e = -1.71 ± 0.04, consistent with exponents estimated in previous studies 1.0 ≤ e ≤ 1.9 and 1.5 ≤ e ≤ 1.9 for soil landslides (Brunetti et al., 2009; Malamud et al., 2004). Further analyses, similar to ours are \n475 \nnecessary to get a better handle on the landslide volume distribution, a critical information with respect to risk analysis and \nlandslide erosion calculation. 5.2.3 Rollover in the distribution of landslide area This represents 92% of total landslides that would \nnot be considered. While the under detection of small landslides would not greatly affect the total landslide volume estimation, \nit could have consequences for our understanding of natural hazards mechanics. This promising result possibly highlights the \nadvantages of using LiDAR data combined to our 3D differencing workflow with low level of change detection to generate \n495 \nmore accurate and complete landslide inventory datasets. 5.2.2 Distribution of landslide volume and area: power-law behaviour ≤ e ≤ 1.9 and 1.5 ≤ e ≤ 1.9 for soil landslides (Brunetti et al., 2009; Malamud et al., 2004). Further analyses, similar to ours are \n475 \nnecessary to get a better handle on the landslide volume distribution, a critical information with respect to risk analysis and \nlandslide erosion calculation. 20 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 5.2.3 Rollover in the distribution of landslide area Several hypothesis have been put forward to explain the rollover behaviour at low landslide area, including the transition to a \ncohesion dominated regime reducing the likelihood of rupture (Frattini and Crosta, 2013; Jeandet et al., 2019; Stark and \n480 \nGuzzetti, 2009), a cohesion gradient with depth (Frattini and Crosta, 2013), landslide amalgamation (Tanyas et al., 2019) or \nthe under detection of small landslides (Hovius et al., 1997; Stark and Hovius, 2001). The landslide area distribution in our \nstudy does not display a rollover. We have proposed a method to rigorously evaluate the likelihood of detecting spurious \nlandslides due to different data sampling of a rough surface at small landslide areas (Fig S.2.2.1). We set a conservative lower bound of 20 m² above which we are confident that all detected changes are true topographic change. We have also checked \n485 \nthat the segmentation distance does not impact the absence of a rollover in our data (Fig S2.2.2 in the supplementary material). Given that Massey et al. (2020) reports a rollover at ~ 50 m² for the Kaikoura earthquake landslide inventory based on 2D \noptical image analysis (Fig.5.a), our results supports the idea that the rollover behaviour observed in previous studies is likely \ncaused by an under detection of small landslides, even with high resolution imagery (Hovius et al., 1997; Stark and Hovius, 2001). If this hypothesis is correct, the number of landslides potentially missed in previous studies can be important. If we \n490 \nconsider the power law fitting statistics for landslide area distribution from Massey et al. (2020), the number of landslides \nbetween 20 m² ≤ A < 500 m² potentially missed would be around 169,000. This represents 92% of total landslides that would \nnot be considered. While the under detection of small landslides would not greatly affect the total landslide volume estimation, \nit could have consequences for our understanding of natural hazards mechanics. This promising result possibly highlights the \nadvantages of using LiDAR data combined to our 3D differencing workflow with low level of change detection to generate \n495 2001). If this hypothesis is correct, the number of landslides potentially missed in previous studies can be important. If we \n490 \nconsider the power law fitting statistics for landslide area distribution from Massey et al. (2020), the number of landslides \nbetween 20 m² ≤ A < 500 m² potentially missed would be around 169,000. 5.2.4 Landslide volume-area relationship The landslide volume-area (V-A) scaling relationships obtained in this study are close to the one of Larsen et al. (2010) for \nsoil landslides. This is consistent with the fact that 50% of the landslide thicknesses are lower than 1 m, showing that most of The landslide volume-area (V-A) scaling relationships obtained in this study are close to the one of Larsen et al. (2010) for \nsoil landslides. This is consistent with the fact that 50% of the landslide thicknesses are lower than 1 m, showing that most of \nour inventory is relevant to shallow landsliding. The V-A scaling relationship of Massey et al. (2020) for soil landslides gives \n500 \nthe best estimation of total landslide volume but overestimate the volume of small landslides. The differences in total volume \npredicted by our two V-A scaling relationships show that estimates of landslide volume deduced from such relationships \ngreatly depend on the method used to fit data. Our results suggest that fitting model on log-binned data gives a better result in \ntotal landslide volume estimation. However, measuring the volume directly from topographic data overcome the issue of \nchoosing a peculiar V-A relationship. 505 3 Toward a limitation of amalgamation and reactivation biases on landslide volume estimatio Hence, \n520 \nour 3D approach detects landslide occurring in steep areas with poor vegetation cover (Fig.6.b) that would have otherwise \nbeen missed with 2D optical imagery approaches, or incorrectly detected with vertical differencing. In this study area, the \nproportion of reactivated landslide area, 27.2%, is lower than new landslide failures (Table.3), and most of the large reactivated \nlandslides that we find have not been included in the initial mapping of 27 landslides by Massey et al. (2018). Assuming that perpendicular to the local topographic slope, as opposed to vertical differencing, is critical in detecting subtle changes. Hence, \n520 \nour 3D approach detects landslide occurring in steep areas with poor vegetation cover (Fig.6.b) that would have otherwise \nbeen missed with 2D optical imagery approaches, or incorrectly detected with vertical differencing. In this study area, the \nproportion of reactivated landslide area, 27.2%, is lower than new landslide failures (Table.3), and most of the large reactivated \nlandslides that we find have not been included in the initial mapping of 27 landslides by Massey et al. (2018). Assuming that perpendicular to the local topographic slope, as opposed to vertical differencing, is critical in detecting subtle changes. Hence, \n520 \nour 3D approach detects landslide occurring in steep areas with poor vegetation cover (Fig.6.b) that would have otherwise \nbeen missed with 2D optical imagery approaches, or incorrectly detected with vertical differencing. In this study area, the \nproportion of reactivated landslide area, 27.2%, is lower than new landslide failures (Table.3), and most of the large reactivated \nlandslides that we find have not been included in the initial mapping of 27 landslides by Massey et al. (2018). Assuming that reactivated landslides cannot be detected by classical methods, the volume potentially lacking represents 29.9 ± 12.8% of the \n525 \ntotal estimated landslide volume, including some of the largest failures of the studied area. Our study area was chosen based \non LiDAR data availability, and may contain a particularly high proportion of reactivated landslides due to the presence of \nactively eroding bare bedrock hillslopes compared to the size of the study area. We expect this proportion to significantly vary \nwhen considering other landscapes with potentially varying proportion of vegetation cover, vegetation density and type (e.g. reactivated landslides cannot be detected by classical methods, the volume potentially lacking represents 29.9 ± 12.8% of the \n525 \ntotal estimated landslide volume, including some of the largest failures of the studied area. 3 Toward a limitation of amalgamation and reactivation biases on landslide volume estimatio The amalgamation effect is a classical issue for 2D landslide mapping and volume assessment (e.g., Li et al., 2014; Marc and \nHovius, 2015b), which leads to a higher number of large landslides than should be expected, and a significant overestimation 21 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. of the total volume of landslides when using a non-linear V-A relationship. A simple solution to this last problem consists in \ndirectly measuring the total landslide volume by comparing pre- and post-earthquake topographic data (e.g., LiDAR) with a \n510 \nDoD (Massey et al., 2020) or with the method described in this paper. While our simple segmentation approach cannot resolve \nthe amalgamation of individual landslides perfectly, the total volume associated to sources and deposit can be robustly \nestimated independantly of a V-A relationship. The detection of reactivated landslides from 2D optical imagery remain challenging due to to the weak contrast of vegetation The detection of reactivated landslides from 2D optical imagery remain challenging due to to the weak contrast of vegetation \nbetween pre- and post-earthquake periods, or even the absence of vegetation. Recent methods have been developped to detect \n515 \nreactivated landslide based on NDVI-trajectory induced by differences in revegetation rate (Behling et al., 2014, 2016) but the \nimpact of possibly missed reactivated landslide on total volume estimation is still poorly understood (Guzzetti et al., 2009). 2D or 3D topographic differencing methods are insensitive to the lack of texture variation and can resolve the issue of \nreactivation. Because vegetation barely develops on very steep slopes and cliffs, the 3D differencing approach oriented between pre- and post-earthquake periods, or even the absence of vegetation. Recent methods have been developped to detect \n515 \nreactivated landslide based on NDVI-trajectory induced by differences in revegetation rate (Behling et al., 2014, 2016) but the \nimpact of possibly missed reactivated landslide on total volume estimation is still poorly understood (Guzzetti et al., 2009). 2D or 3D topographic differencing methods are insensitive to the lack of texture variation and can resolve the issue of \nreactivation. Because vegetation barely develops on very steep slopes and cliffs, the 3D differencing approach oriented g\ny\np\ny\np\np\ng\npp\nperpendicular to the local topographic slope, as opposed to vertical differencing, is critical in detecting subtle changes. 3 Toward a limitation of amalgamation and reactivation biases on landslide volume estimatio Our study area was chosen based \non LiDAR data availability, and may contain a particularly high proportion of reactivated landslides due to the presence of \nactively eroding bare bedrock hillslopes compared to the size of the study area. We expect this proportion to significantly vary \nwhen considering other landscapes with potentially varying proportion of vegetation cover, vegetation density and type (e.g. grass, shrubs, trees), lithology and ground shaking intensity. Nonetheless, our finding represents a first approach to the issue \n530 \nof considering reactivated landslides in total landslide volume estimates, and our results indicates that this should not be \nneglected at least in regions dominated by a low or absent vegetation cover. To evaluate the difference of volume that would have been estimated from traditional methods impacted by under-detection \nof reactivated landslides, we apply the Massey et al. (2020) V-A relationship for all landslides only to the new failures detected grass, shrubs, trees), lithology and ground shaking intensity. Nonetheless, our finding represents a first approach to the issue \n530 \nof considering reactivated landslides in total landslide volume estimates, and our results indicates that this should not be \nneglected at least in regions dominated by a low or absent vegetation cover. To evaluate the difference of volume that would have been estimated from traditional methods impacted by under-detection \nof reactivated landslides, we apply the Massey et al. (2020) V-A relationship for all landslides only to the new failures detected neglected at least in regions dominated by a low or absent vegetation cover. To evaluate the difference of volume that would have been estimated from traditional methods impacted by under-detection \nof reactivated landslides, we apply the Massey et al. (2020) V-A relationship for all landslides only to the new failures detected in our inventory. As our inventory has very likely a much lower detection level than optical methods (see 5.2.3), we only \n535 \nconsider new failures with a minimum area of 50 m² (typical of the rollover observed in Massey et al. 2020). This amount to \na 2D traditional processing of the study area. This traditional approach would predict a total volume of 616,308 m3, \nsignificantly lower than the volume measured directly by our approach. This highlights the need to more systematically \ngeneralize the use of 3D data to improve the creation of robust landslide inventories and generate more accurate estimate of \nthe total volume of sediment produced by earthquakes or climatic events. 6. Conclusion In this paper, we introduced a new workflow for semi-automated landslide detection and volume estimation using 3D \ndifferencing based on high resolution topographic point cloud data. This method uses the M3C2 algorithm developed by Lague \net al. (2013) for accurate change detection based on the 3D distance normal to the local surface as well as a vertical-M3C2 for \nvolume calculation of landslide sources and deposits once a 3D connected component segmentation procedure has been applied \n545 \nto individualize landslides. Spatially variable uncertainties on distance and volume are provided by the calculation and used \nin the workflow to evaluate if a change is statistically significant or not, and for volume uncertainty estimation. We provide \nvarious tests and recipes to estimate the registration error and choose the parameters of the M3C2 algorithm as function of the \npoint cloud density to ensure the lowest level of change detection, and the best resolution of the 3D map of change. Applied to a 5 km² area located in the Kaikoura region in New Zealand with pre- and post-earthquake LiDAR, we showed that: \n550 a 5 km² area located in the Kaikoura region in New Zealand with pre- and post-earthquake LiDAR  \nA level of 3D change detection at 95% confidence of 0.34 m can be reached with airborne LiDAR data, which is \nlargely set by the registration error. Because it operates on raw data, M3C2 accounts for sub-pixel characteristics such \nas point density and roughness that are not accounted for when working on DEMs, and results in more robust statistics \nwhen it comes to evaluate if a change is significant or not. 3D point cloud differencing is critical on steep slopes, and \nallows to decrease the level of change detection compared to the traditional DoD.  \nAdding elevation information in landslide detection removes the amalgamation effect on the total landslide volume \nestimation by directly measuring it in 3D rather than considering an ad hoc V-A relationship. Amalgamation in 3D \nis still a potential issue when exploring individual landslide area and volume statistics given the simplistic \nsegmentation approach that we have used, however our approach has the benefits of more systematically capturing \nsmall landslides than traditional approaches based on 2D imagery with manual or automatic landslide mapping.  \nLandslide reactivation on surfaces lacking a significant vegetation cover is classically missed with 2D imagery \nprocessing due to the lack of texture or spectral change. 3 Toward a limitation of amalgamation and reactivation biases on landslide volume estimatio 540 in our inventory. As our inventory has very likely a much lower detection level than optical methods (see 5.2.3), we only \n535 \nconsider new failures with a minimum area of 50 m² (typical of the rollover observed in Massey et al. 2020). This amount to \na 2D traditional processing of the study area. This traditional approach would predict a total volume of 616,308 m3, \nsignificantly lower than the volume measured directly by our approach. This highlights the need to more systematically \ngeneralize the use of 3D data to improve the creation of robust landslide inventories and generate more accurate estimate of \nthe total volume of sediment produced by earthquakes or climatic events. 540 22 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Author contribution D.L. and T.B. designed the landslide detection workflow. T.B. and all co-authors participated to the discussion, writing and \nreviewing of this paper. T.B produced the figures and the code. Supplement link The supplement related to this article is available online at: . 590 The supplement related to this article is available online at: . 590 6. Conclusion Current bottlenecks to apply this \nworkflow over larger scales, beyond the availability of 3D data itself, are the registration of pre- and post-EQ data when \ncomplex co-seismic deformation patterns occur, and limitations of the segmentation method in high landslide density areas. While airborne LiDAR is best suited to vegetated environments and currently results in the best precision compared to aerial \n580 \nor spatial photogrammetry, the workflow operates for any kind of 3D data. While airborne LiDAR is best suited to vegetated environments and currently results in the best precision compared to aerial \n580 \nor spatial photogrammetry, the workflow operates for any kind of 3D data. While airborne LiDAR is best suited to vegetated environments and currently results in the best precision compared to aerial \n580 \nor spatial photogrammetry, the workflow operates for any kind of 3D data. Code availability The code producing the landslide inventory in this study is available as a jupyter notebook form at https://github.com/Thomas-\nBrd/3D_landslide_detection, and is also archived in Zenodo: http://doi.org/10.5281/zenodo.4010806. Data availability \n585 \nLiDAR data used in this study can be found at: http://doi.org/10.5281/zenodo.4011629 \nThe final landslide source and deposit information supporting the findings of this paper can be found at \nhttps://github.com/Thomas-Brd/3D_landslide_detection, or in Zenodo: http://doi.org/10.5281/zenodo.4010806 6. Conclusion 3D processing fully accounts for reactivated landslides. In \nour study area, 29.9 ± 12.8 % of the total volume was due to landslide reactivation, highlighting that in areas with a \nmixture of vegetated and non-vegetated steep slopes, 2D approaches can significantly underestimate the number and \nvolume of landslides.  \nAs this method provides direct 3D measurement, landslide geometry properties can be explored and tested such as \nthe V-A relationship, landslide area and volume distribution and others. Our results are broadly consistent with the \nV-A relationship scaling parameters determined by Larsen et al. (2010) and Massey et al. (2020) for soil landslides, \nwith a scaling exponent of 1.20. The largest and deepest landslides deviate significantly from this trend, but they are \ntoo few in our database to confidently infer a scaling relationship for these.  \nNo rollover is observed in the landslide area distribution down to 20 m², our conservative resolution limit. Inventories \nbased on 2D images analysis following the Kaikoura EQ typically observe a rollover at 50 m² (Massey et al., 2020).  \nNo rollover is observed in the landslide area distribution down to 20 m², our conservative resolution limit. Inventories \nbased on 2D images analysis following the Kaikoura EQ typically observe a rollover at 50 m² (Massey et al., 2020). 23 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. This lend credit to the hypothesis that the rollover systematically observed in landslide area distributions generated \nfrom 2D images is related to an under detection of small landslide. Our 3D processing workflow is a first step towards harnessing the full potential of repeat 3D high resolution topographic \n575 \nsurveys to automatically create complete and accurate landslide inventories that are critically needed to improve landslide \nscience and managing the cascade of hazards following large earthquakes or storm events. Current bottlenecks to apply this \nworkflow over larger scales, beyond the availability of 3D data itself, are the registration of pre- and post-EQ data when \ncomplex co-seismic deformation patterns occur, and limitations of the segmentation method in high landslide density areas. Our 3D processing workflow is a first step towards harnessing the full potential of repeat 3D high resolution topographic \n575 \nsurveys to automatically create complete and accurate landslide inventories that are critically needed to improve landslide \nscience and managing the cascade of hazards following large earthquakes or storm events. References \n600 Aerial Surveys, Aerial photographs derived from two surveys of the study area carried out in 2014 to 2015 and in 2016 to \n2017, by Aerial Surveys Ltd, 2017. Behling, R., Roessner, S., Kaufmann, H. and Kleinschmit, B.: Automated spatiotemporal landslide mapping over large areas \nusing rapideye time series data, Remote Sens., 6(9), 8026–8055, doi:10.3390/rs6098026, 2014. Aerial Surveys, Aerial photographs derived from two surveys of the study area carried out in 2014 to 2015 and in 2016 to \n2017, by Aerial Surveys Ltd, 2017. Behling, R., Roessner, S., Kaufmann, H. and Kleinschmit, B.: Automated spatiotemporal landslide mapping over large areas \nusing rapideye time series data, Remote Sens., 6(9), 8026–8055, doi:10.3390/rs6098026, 2014. Behling, R., Roessner, S., Kaufmann, H. and Kleinschmit, B.: Automated spatiotemporal landslide mapping over large areas \nusing rapideye time series data, Remote Sens., 6(9), 8026–8055, doi:10.3390/rs6098026, 2014. Behling, R., Roessner, S., Golovko, D. and Kleinschmit, B.: Derivation of long-term spatiotemporal landslide activity—A \n605 \nmulti-sensor time series approach, Remote Sens. Environ., 186, 88–104, doi:10.1016/j.rse.2016.07.017, 2016. Behling, R., Roessner, S., Golovko, D. and Kleinschmit, B.: Derivation of long-term spatiotemporal landslide activity—A \n605 \nmulti-sensor time series approach, Remote Sens. Environ., 186, 88–104, doi:10.1016/j.rse.2016.07.017, 2016. Besl, P. J. and McKay, N. D.: A Method for Registration of 3-D Shapes, IEEE Trans. Pattern Anal. Mach. Intell., 14(2), 239– Behling, R., Roessner, S., Golovko, D. and Kleinschmit, B.: Derivation of long-term spatiotemporal landslide activity—A \n605 \nmulti-sensor time series approach, Remote Sens. Environ., 186, 88–104, doi:10.1016/j.rse.2016.07.017, 2016. Besl, P. J. and McKay, N. D.: A Method for Registration of 3-D Shapes, IEEE Trans. Pattern Anal. Mach. Intell., 14(2), 239–\n256, doi:10.1109/34.121791, 1992. Bossi, G., Cavalli, M., Crema, S., Frigerio, S., Quan Luna, B., Mantovani, M., Marcato, G., Schenato, L. and Pasuto, A.: Multi- e\ng,\n.,\noess e , S., Go ov o,\n. a d\ne sc\nt,\n.:\ne vat o\no\no g te\nspat ote po a\na ds de act v ty\n605\nmulti-sensor time series approach, Remote Sens. Environ., 186, 88–104, doi:10.1016/j.rse.2016.07.017, 2016. Besl, P. J. and McKay, N. D.: A Method for Registration of 3-D Shapes, IEEE Trans. Pattern Anal. Mach. Intell., 14(2), 239–\n256, doi:10.1109/34.121791, 1992. Besl, P. J. and McKay, N. D.: A Method for Registration of 3-D Shapes, IEEE Trans. Pattern Anal. Mach. Intell., 14(2), 239–\n256, doi:10.1109/34.121791, 1992. Competing interests The authors declare that they have no conflict of interest. 595 24 24 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Acknowledgements This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 \nresearch and innovation programme (grant agreement no. 803721), from the Agence Nationale de la Recherche (grant no. ANR-14-CE33-0005), from the LiDAR topobathymetric platform and from the Brittany region. References \n600 Lett., 46(8), 4258–4266, doi:10.1029/2019GL082351, 2019. Keefer, D. K.: The importance of earthquake-induced landslides to long-term slope erosion and slope-failure hazards in \nseismically active regions, Geomorphology, 10(1–4), 265–284, doi:10.1016/0169-555X(94)90021-3, 1994. Lague, D., Brodu, N. and Leroux, J.: Accurate 3D comparison of complex topography with terrestrial laser scanner: \n645 \nApplication \nto \nthe \nRangitikei \ncanyon \n(N-Z), \nISPRS \nJ. Photogramm. Remote \nSens., \n82, \n10–26, \ndoi:10.1016/j.isprsjprs.2013.04.009, 2013. Larsen, I. J., Montgomery, D. R. and Korup, O.: Landslide erosion controlled by hillslope material, Nat. Geosci., 3(4), 247–\n251, doi:10.1038/ngeo776, 2010. Lague, D., Brodu, N. and Leroux, J.: Accurate 3D comparison of complex topography with terrestrial laser scanner: \n645 \nApplication \nto \nthe \nRangitikei \ncanyon \n(N-Z), \nISPRS \nJ. Photogramm. Remote \nSens., \n82, \n10–26, \ndoi:10.1016/j.isprsjprs.2013.04.009, 2013. Larsen I J Montgomery D R and Korup O : Landslide erosion controlled by hillslope material Nat Geosci\n3(4) 247 Larsen, I. J., Montgomery, D. R. and Korup, O.: Landslide erosion controlled by hillslope material, Nat. Geosci., 3(4), 247–\n251, doi:10.1038/ngeo776, 2010. Li, G., West, A. J., Densmore, A. L., Jin, Z., Parker, R. N. and Hilton, R. G.: Seismic mountain building: Landslides associated \n650 \nwith the 2008 Wenchuan earthquake in the context of a generalized model for earthquake volume balance, Geochemistry, \nGeophys. Geosystems, 15(4), 833–844, doi:10.1002/2013GC005067, 2014. Li, G., West, A. J., Densmore, A. L., Jin, Z., Parker, R. N. and Hilton, R. G.: Seismic mountain building: Landslides associated \n650 \nwith the 2008 Wenchuan earthquake in the context of a generalized model for earthquake volume balance, Geochemistry, \nGeophys. Geosystems, 15(4), 833–844, doi:10.1002/2013GC005067, 2014. Lumia, R., Shapiro, L. and Zuniga, O.: A new connected components algorithm for virtual memory computers, Comput. Vision, Graph. Image Process., 22(2), 287–300, doi:10.1016/0734-189X(83)90071-3, 1983. Lumia, R., Shapiro, L. and Zuniga, O.: A new connected components algorithm for virtual memory computers, Comput. Vision, Graph. Image Process., 22(2), 287–300, doi:10.1016/0734-189X(83)90071-3, 1983. Malamud, B. D. and Turcotte, D. L.: Self-organized criticality applied to natural hazards, Nat. Hazards, 20(2–3), 93–116, \n655 \ndoi:10.1023/A:1008014000515, 1999. Malamud, B. D., Turcotte, D. L., Guzzetti, F. and Reichenbach, P.: Landslide inventories and their statistical properties, Earth \nSurf. Process. Landforms, 29(6), 687–711, doi:10.1002/esp.1064, 2004. Marc, O. and Hovius, N.: Amalgamation in landslide maps: Effects and automatic detection, Nat. Hazards Earth Syst. Sci., Malamud, B. D. and Turcotte, D. L.: Self-organized criticality applied to natural hazards, Nat. Hazards, 20(2–3), 93–116, \n655 \ndoi:10.1023/A:1008014000515, 1999. Malamud, B. D., Turcotte, D. L., Guzzetti, F. References \n600 temporal LiDAR-DTMs as a tool for modelling a complex landslide: A case study in the Rotolon catchment (eastern Italian \n610 \nAlps), Nat. Hazards Earth Syst. Sci., 15(4), 715–722, doi:10.5194/nhess-15-715-2015, 2015. Bull, J. M., Miller, H., Gravley, D. M., Costello, D., Hikuroa, D. C. H. and Dix, J. K.: Assessing debris flows using LIDAR \ndifferencing: \n18 \nMay \n2005 \nMatata \nevent, \nNew \nZealand, \nGeomorphology, \n124(1–2), \n75–84, \ndoi:10.1016/j.geomorph.2010.08.011, 2010. Corsini, A., Borgatti, L., Cervi, F., Dahne, A., Ronchetti, F. and Sterzai, P.: Estimating mass-wasting processes in active earth \n615 \nslides - Earth flows with time-series of High-Resolution DEMs from photogrammetry and airborne LiDAR, Nat. Hazards \nEarth Syst. Sci., 9(2), 433–439, doi:10.5194/nhess-9-433-2009, 2009. Dolan J.F, Rhodes E.J.. Marlborough Fault System, South Island, New Zealand, airborne lidar. National Center for Airborne \nLaser Mapping (NCALM), distributed by OpenTopography. http://dx.doi.org/10.5069/G9G44N75, 2016. Dolan J.F, Rhodes E.J.. Marlborough Fault System, South Island, New Zealand, airborne lidar. National Center for Airborne \nLaser Mapping (NCALM), distributed by OpenTopography. http://dx.doi.org/10.5069/G9G44N75, 2016. Esposito, G., Salvini, R., Matano, F., Sacchi, M., Danzi, M., Somma, R. and Troise, C.: Multitemporal monitoring of a coastal \n620 \nlandslide through SfM-derived point cloud comparison, Photogramm. Rec., 32(160), 459–479, doi:10.1111/phor.12218, 2017. Frattini, P. and Crosta, G. B.: The role of material properties and landscape morphology on landslide size distributions, Earth \nPlanet. Sci. Lett., 361, 310–319, doi:10.1016/j.epsl.2012.10.029, 2013. Guzzetti, F.: Geomorphology Morphological and kinematic evolution of a large earth fl ow : The Montaguto landslide , \n625 \nsouthern Italy, Geomorphology, 187, 61–79, doi:10.1016/j.geomorph.2012.12.035, 2013. Guzzetti, F., Malamud, B. D., Turcotte, D. L. and Reichenbach, P.: Power-law correlations of landslide areas in central Italy, 25 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Earth Planet. Sci. Lett., 195(3–4), 169–183, doi:10.1016/S0012-821X(01)00589-1, 2002. Guzzetti, F., Ardizzone, F., Cardinali, M., Rossi, M. and Valigi, D.: Landslide volumes and landslide mobilization rates in Umbria, central Italy, Earth Planet. Sci. Lett., 279(3–4), 222–229, doi:10.1016/j.epsl.2009.01.005, 2009. 630 \nGuzzetti, F., Mondini, A. C., Cardinali, M., Fiorucci, F., Santangelo, M. and Chang, K. T.: Landslide inventory maps: New \ntools for an old problem, Earth-Science Rev., 112(1–2), 42–66, doi:10.1016/j.earscirev.2012.02.001, 2012. Hovius, N., Stark, C. P. and Allen, P. A.: Sediment flux from a mountain belt derived by landslide mapping, Geology, 25(3), \n231, doi:10.1130/0091-7613(1997)025<0231:SFFAMB>2.3.CO;2, 1997. Guzzetti, F., Mondini, A. C., Cardinali, M., Fiorucci, F., Santangelo, M. and Chang, K. References \n600 T.: Landslide inventory maps: New \ntools for an old problem, Earth-Science Rev., 112(1–2), 42–66, doi:10.1016/j.earscirev.2012.02.001, 2012. Hovius, N., Stark, C. P. and Allen, P. A.: Sediment flux from a mountain belt derived by landslide mapping, Geology, 25(3), \n231, doi:10.1130/0091-7613(1997)025<0231:SFFAMB>2.3.CO;2, 1997. Hovius, N., Stark, C. P. and Allen, P. A.: Sediment flux from a mountain belt derived by landslide mapping, Geology, 25(3), \n231, doi:10.1130/0091-7613(1997)025<0231:SFFAMB>2.3.CO;2, 1997. Hovius, N., Stark, C. P. and Allen, P. A.: Sediment flux from a mountain belt derived by landslide mapping, Geology, 25(3), \n231, doi:10.1130/0091-7613(1997)025<0231:SFFAMB>2.3.CO;2, 1997. James, M. R., Robson, S. and Smith, M. W.: 3-D uncertainty-based topographic change detection with structure-from-motion \n635 \nphotogrammetry: precision maps for ground control and directly georeferenced surveys, Earth Surf. Process. Landforms, \n42(12), 1769–1788, doi:10.1002/esp.4125, 2017. James, M. R., Robson, S. and Smith, M. W.: 3-D uncertainty-based topographic change detection with structure-from-motion \n635 \nphotogrammetry: precision maps for ground control and directly georeferenced surveys, Earth Surf. Process. Landforms, \n42(12), 1769–1788, doi:10.1002/esp.4125, 2017. James, P. I., Ph, D., Dolan, J. F. and Hall, Z.: Data Collection and Processing Report LiDAR survey of five fault segments ( \nEastern Clarence , Western Clarence , Central Eastern Awatere , West Wairau and East Hope-Conway ) of the Marlborough James, P. I., Ph, D., Dolan, J. F. and Hall, Z.: Data Collection and Processing Report LiDAR survey of five fault segments ( \nEastern Clarence , Western Clarence , Central Eastern Awatere , West Wairau and East Hope-Conway ) of the Marlborough \nFault System on the Northwestern portion of New Zealand ’ s S, , (213), n.d. 640 Fault System on the Northwestern portion of New Zealand ’ s S, , (213), n.d. 640 \nJeandet, L., Steer, P., Lague, D. and Davy, P.: Coulomb Mechanics and Relief Constraints Explain Landslide Size Distribution, \nGeophys. Res. Lett., 46(8), 4258–4266, doi:10.1029/2019GL082351, 2019. Keefer, D. K.: The importance of earthquake-induced landslides to long-term slope erosion and slope-failure hazards in \nseismically active regions, Geomorphology, 10(1–4), 265–284, doi:10.1016/0169-555X(94)90021-3, 1994. Fault System on the Northwestern portion of New Zealand s S, , (213), n.d. 640 \nJeandet, L., Steer, P., Lague, D. and Davy, P.: Coulomb Mechanics and Relief Constraints Explain Landslide Size Distribution, \nGeophys. Res. Lett., 46(8), 4258–4266, doi:10.1029/2019GL082351, 2019. Keefer, D. K.: The importance of earthquake-induced landslides to long-term slope erosion and slope-failure hazards in Jeandet, L., Steer, P., Lague, D. and Davy, P.: Coulomb Mechanics and Relief Constraints Explain Landslide Size Distribution, \nGeophys. Res. References \n600 and Reichenbach, P.: Landslide inventories and their statistical properties, Earth \nSurf. Process. Landforms, 29(6), 687–711, doi:10.1002/esp.1064, 2004. Surf. Process. Landforms, 29(6), 687–711, doi:10.1002/esp.1064, 2004. Marc, O. and Hovius, N.: Amalgamation in landslide maps: Effects and automatic detection, Nat. Hazards Earth Syst. Sci., Marc, O. and Hovius, N.: Amalgamation in landslide maps: Effects and automatic detection, Nat. Hazards Earth Syst. Sci., \n15(4), 723–733, doi:10.5194/nhess-15-723-2015, 2015a. 660 \nM\nO\nd H\ni\nN\nA\nl\ni\ni\nl\nd lid\nEff\nd\ni d\ni\nN\nH\nd E\nh S\nS i 15(4), 723–733, doi:10.5194/nhess-15-723-2015, 2015a. 660 \nMarc, O. and Hovius, N.: Amalgamation in landslide maps: Effects and automatic detection, Nat. Hazards Earth Syst. Sci., Marc, O. and Hovius, N.: Amalgamation in landslide maps: Effects and automatic detection, Nat. Hazards Earth Syst. Sci., 26 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. 15(4), 723–733, doi:10.5194/nhess-15-723-2015, 2015b. Marc, O., Hovius, N. and Meunier, P.: The mass balance of earthquakes and earthquake sequences, Geophys. Res. Lett., 43(8), \n3708–3716, doi:10.1002/2016GL068333, 2016. 3708–3716, doi:10.1002/2016GL068333, 2016. Marc, O., Behling, R., Andermann, C., Turowski, J. M., Illien, L., Roessner, S. and Hovius, N.: Long-term erosion of the \n665 \nNepal Himalayas by bedrock landsliding: The role of monsoons, earthquakes and giant landslides, Earth Surf. Dyn., 7(1), 107–\n128, doi:10.5194/esurf-7-107-2019, 2019. Martha, T. R., Kerle, N., Jetten, V., van Westen, C. J. and Kumar, K. V.: Characterising spectral, spatial and morphometric \nproperties of landslides for semi-automatic detection using object-oriented methods, Geomorphology, 116(1–2), 24–36, 665 Marc, O., Behling, R., Andermann, C., Turowski, J. M., Illien, L., Roessner, S. and Hovius, N.: Long-term erosion of the \n665 \nNepal Himalayas by bedrock landsliding: The role of monsoons, earthquakes and giant landslides, Earth Surf. Dyn., 7(1), 107–\n128, doi:10.5194/esurf-7-107-2019, 2019. Martha, T. R., Kerle, N., Jetten, V., van Westen, C. J. and Kumar, K. V.: Characterising spectral, spatial and morphometric doi:10.1016/j.geomorph.2009.10.004, 2010. 670 \nMassey, C., Townsend, D., Rathje, E., Allstadt, K. E., Lukovic, B., Kaneko, Y., Bradley, B., Wartman, J., Jibson, R. W., \nPetley, D. N., Horspool, N., Hamling, I., Carey, J., Cox, S., Davidson, J., Dellow, S., Godt, J. W., Holden, C., Jones, K., Kaiser, \nA., Little, M., Lyndsell, B., McColl, S., Morgenstern, R., Rengers, F. K., Rhoades, D., Rosser, B., Strong, D., Singeisen, C. References \n600 and Villeneuve, M.: Landslides triggered by the 14 November 2016 Mw 7.8 Kaikōura earthquake, New Zealand, Bull. Seismol. Soc. Am., 108(3B), 1630–1648, doi:10.1785/0120170305, 2018. 675 \nMassey, C. I., Townsend, D., Jones, K., Lukovic, B., Rhoades, D., Morgenstern, R., Rosser, B., Ries, W., Howarth, J., Hamling, \nI., Petley, D., Clark, M., Wartman, J., Litchfield, N. and Olsen, M.: Volume characteristics of landslides triggered by the M \nW 7.8 2016 Kaikōura Earthquake, New Zealand, derived from digital surface difference modelling , J. Geophys. Res. Earth \nSurf., 0–3, doi:10.1029/2019jf005163, 2020. Mora, O. E., Gabriela Lenzano, M., Toth, C. K., Grejner-Brzezinska, D. A. and Fayne, J. V.: Landslide change detection based \n680 \non Multi-Temporal airborne LIDAR-derived DEMs, Geosci., 8(1), 6–8, doi:10.3390/geosciences8010023, 2018. Mouyen, M., Steer, P., Chang, K.-J., Le Moigne, N., Hwang, C., Hsieh, W.-C., Jeandet, L., Longuevergne, L., Cheng, C.-C., \nBoy, J.-P. and Masson, F.: Quantifying sediment mass redistribution from joint time-lapse gravimetry and photogrammetry \nsurveys, Earth Surf. Dyn. Discuss., (July), 1–37, doi:10.5194/esurf-2019-35, 2019. Mora, O. E., Gabriela Lenzano, M., Toth, C. K., Grejner-Brzezinska, D. A. and Fayne, J. V.: Landslide change detection based \n680 \non Multi-Temporal airborne LIDAR-derived DEMs, Geosci., 8(1), 6–8, doi:10.3390/geosciences8010023, 2018. Parker, R. N., Densmore, A. L., Rosser, N. J., De Michele, M., Li, Y., Huang, R., Whadcoat, S. and Petley, D. N.: Mass wasting \n685 \ntriggered by the 2008 Wenchuan earthquake is greater than orogenic growth, Nat. Geosci., 4(7), 449–452, \ndoi:10.1038/ngeo1154, 2011. Passalacqua, P., Belmont, P., Staley, D. M., Simley, J. D., Arrowsmith, J. R., Bode, C. A., Crosby, C., DeLong, S. B., Glenn, \nN. F., Kelly, S. A., Lague, D., Sangireddy, H., Schaffrath, K., Tarboton, D. G., Wasklewicz, T. and Wheaton, J. M.: Analyzing Parker, R. N., Densmore, A. L., Rosser, N. J., De Michele, M., Li, Y., Huang, R., Whadcoat, S. and Petley, D. N.: Mass wasting \n685 \ntriggered by the 2008 Wenchuan earthquake is greater than orogenic growth, Nat. Geosci., 4(7), 449–452, \ndoi:10.1038/ngeo1154, 2011. P\nl\nP B l\nt P St l\nD M Si l\nJ D A\nith J R B d\nC A C\nb\nC D L\nS B Gl Passalacqua, P., Belmont, P., Staley, D. M., Simley, J. D., Arrowsmith, J. R., Bode, C. A., Crosby, C., DeLong, S. temporal airborne LiDAR data: The Montaguto landslide (Southern Italy), Remote Sens. Environ., 115(12), 3237–3248, \n705 \ndoi:10.1016/j.rse.2011.07.007, 2011. \nWheaton, J. M., Brasington, J., Darby, S. E. and Sear, D. A.: Accounting for uncertainty in DEMs from repeat topographic \nsurveys: Improved sediment budgets, Earth Surf. Process. Landforms, 35(2), 136–156, doi:10.1002/esp.1886, 2010. References \n600 B., Glenn, \nN F Kelly S A Lague D Sangireddy H Schaffrath K Tarboton D G Wasklewicz T and Wheaton J M : Analyzing high resolution topography for advancing the understanding of mass and energy transfer through landscapes: A review, Earth-\n690 \nScience Rev., 148, 174–193, doi:10.1016/j.earscirev.2015.05.012, 2015. Pradhan, B., Jebur, M. N., Shafri, H. Z. M. and Tehrany, M. S.: Data fusion technique using wavelet transform and taguchi \nmethods for automatic landslide detection from airborne laser scanning data and quickbird satellite imagery, IEEE Trans. Geosci. Remote Sens., 54(3), 1610–1622, doi:10.1109/TGRS.2015.2484325, 2016. Stark, C. P. and Guzzetti, F.: Landslide rupture and the probability distribution of mobilized debris volumes, J. Geophys. Res. 695 27 https://doi.org/10.5194/esurf-2020-73\nPreprint. Discussion started: 15 September 2020\nc⃝Author(s) 2020. CC BY 4.0 License. Earth Surf., 114(2), 1–16, doi:10.1029/2008JF001008, 2009. Stumpf, A., Malet, J. P., Allemand, P. and Ulrich, P.: Surface reconstruction and landslide displacement measurements with \nPléiades satellite images, ISPRS J. Photogramm. Remote Sens., 95, 1–12, doi:10.1016/j.isprsjprs.2014.05.008, 2014. Stumpf, A., Malet, J. P., Allemand, P., Pierrot-Deseilligny, M. and Skupinski, G.: Ground-based multi-view photogrammetry \nfor \nthe \nmonitoring \nof \nlandslide \ndeformation \nand \nerosion, \nGeomorphology, \n231(October), \n130–145, \n700 \ndoi:10.1016/j.geomorph.2014.10.039, 2015. Tanyaş, H., van Westen, C. J., Allstadt, K. E. and Jibson, R. W.: Factors controlling landslide frequency–area distributions, \nEarth Surf. Process. Landforms, 44(4), 900–917, doi:10.1002/esp.4543, 2019. Ventura, G., Vilardo, G., Terranova, C. and Sessa, E. B.: Tracking and evolution of complex active landslides by multi- for \nthe \nmonitoring \nof \nlandslide \ndeformation \nand \nerosion, \nGeomorphology, \n231(October), \n130–145, \n700 \ndoi:10.1016/j.geomorph.2014.10.039, 2015. Tanyaş, H., van Westen, C. J., Allstadt, K. E. and Jibson, R. W.: Factors controlling landslide frequency–area distributions, \nEarth Surf. Process. Landforms, 44(4), 900–917, doi:10.1002/esp.4543, 2019. Ventura, G., Vilardo, G., Terranova, C. and Sessa, E. B.: Tracking and evolution of complex active landslides by multi- Tanyaş, H., van Westen, C. J., Allstadt, K. E. and Jibson, R. W.: Factors controlling landslide frequency–area distributions, \nEarth Surf. Process. Landforms, 44(4), 900–917, doi:10.1002/esp.4543, 2019. temporal airborne LiDAR data: The Montaguto landslide (Southern Italy), Remote Sens. Environ., 115(12), 3237–3248, \n705 \ndoi:10.1016/j.rse.2011.07.007, 2011. Wheaton, J. M., Brasington, J., Darby, S. E. and Sear, D. A.: Accounting for uncertainty in DEMs from repeat topographic \nsurveys: Improved sediment budgets, Earth Surf. Process. Landforms, 35(2), 136–156, doi:10.1002/esp.1886, 2010. Wheaton, J. M., Brasington, J., Darby, S. E. and Sear, D. A.: Accounting for uncertainty in DEMs from repeat topographic \nsurveys: Improved sediment budgets, Earth Surf. Process. References \n600 Landforms, 35(2), 136–156, doi:10.1002/esp.1886, 2010. 710 28"