Comparability of multi‐temporal DTMs derived from different LiDAR platforms: Error sources and uncertainties in the application of geomorphic impact studies

Abstract Multi‐temporal digital terrain models (DTMs) derived from airborne or uncrewed aerial vehicle (UAV)‐borne light detection and ranging (LiDAR) platforms are frequently used tools in geomorphic impact studies. Accurate estimation of mobilized sediments from multi‐temporal DTMs is indispensable for hazard assessment. To study volumetric changes in alpine environments it is crucial to identify and discuss different kind of error sources in multi‐temporal data. We subdivided errors into those caused by data acquisition, data processing, and spatial properties of the terrain. In terms of the quantification of surface changes, the propagation of errors can lead to high uncertainties. Three alpine catchments with different LiDAR point clouds of different origins (airborne laser scanning [ALS], UAV‐borne laser scanning [ULS]), varying point densities, accuracies and qualities were analysed, and used as basis for interpolating DTMs. The workflow was developed in the Schöttlbach area in Styria and later applied to further catchments in Austria. The main aim of the presented work is a comprehensive DTM uncertainty analysis specially designed for geomorphic impact studies, with a resulting uncertainty analysis serving as input for a change detection tool. Our findings reveal that geomorphic impact studies need the careful distinction between actual surface changes and different data uncertainties. ULS combines the benefits of terrestrial laser scanning with all the benefits of ALS. However, the use of ULS data does not necessarily improve the results of the analysis since the high level of detail is not always helpful in geomorphic impact studies. In order to make the different point clouds and DTMs comparable the quality of the ULS point cloud had to be reduced to fit the accuracy of the reference data (older ALS point clouds). Using a point cloud with a high point density with a regular planimetric point spacing and less data gaps, in the best case collected during leaf‐off conditions (e.g., cross‐flight strategy) turned out to be sufficient for our geomorphic research purposes.

changes in alpine environments it is crucial to identify and discuss different kind of error sources in multi-temporal data. We subdivided errors into those caused by data acquisition, data processing, and spatial properties of the terrain. In terms of the quantification of surface changes, the propagation of errors can lead to high uncertainties.
Three alpine catchments with different LiDAR point clouds of different origins (airborne laser scanning [ALS], UAV-borne laser scanning [ULS]), varying point densities, accuracies and qualities were analysed, and used as basis for interpolating DTMs.
The workflow was developed in the Schöttlbach area in Styria and later applied to further catchments in Austria. The main aim of the presented work is a comprehensive DTM uncertainty analysis specially designed for geomorphic impact studies, with a resulting uncertainty analysis serving as input for a change detection tool. Our findings reveal that geomorphic impact studies need the careful distinction between actual surface changes and different data uncertainties. ULS combines the benefits of terrestrial laser scanning with all the benefits of ALS. However, the use of ULS data does not necessarily improve the results of the analysis since the high level of detail is not always helpful in geomorphic impact studies. In order to make the different point clouds and DTMs comparable the quality of the ULS point cloud had to be reduced to fit the accuracy of the reference data (older ALS point clouds). Using a point cloud with a high point density with a regular planimetric point spacing and less data gaps, in the best case collected during leaf-off conditions (e.g., cross-flight strategy) turned out to be sufficient for our geomorphic research purposes.

K E Y W O R D S
airborne LiDAR, digital terrain model, DTMs of difference, UAV-borne LiDAR, uncertainty analysis

| INTRODUCTION
Since the beginning of the 21st century, LiDAR (light detection and ranging; also known as laser scanning) has been a frequently used active remote sensing technique for studying geomorphic processes.
LiDAR data increases the level of detail for geomorphological mapping and improves the quality of terrain data (Tarolli et al., 2009). The use of LiDAR data has increased exponentially as high-resolution topographic data are nowadays widely available and accessible.
Because of its ability to penetrate vegetation cover and acquire point data with high density, precision, and accuracy, LiDAR provides an accurate and high-resolution representation of the earth's surface.
New technological advances in geomorphology (Viles, 2016) and the transition from interpolating a digital terrain model (DTM) from a small amount of terrestrial surveyed terrain points to millions of LiDAR ground points improved the quality and reliability of geomorphometric analysis. Hence, high-density LiDAR data has become a powerful asset in earth surface research (Fujii & Fukuchi, 2005;Jaboyedoff & Derron, 2020;Shan & Toth, 2017;Vosselman & Maas, 2010). Several laser scanning systems exist such as terrestrial laser scanning (TLS), airborne laser scanning (ALS), and mobile laser scanning (MLS) systems (e.g., mounted on a car or boat). A rather recent development in LiDAR technology is the application of MLS-systems mounted on light-weight uncrewed aerial vehicles (UAVs; UAV-borne laser scanning [ULS]) for monitoring comparably small study areas (< 5 km 2 ) (Bremer et al., 2019a). The significant advantage of ULS is that the benefits of TLS (e.g., short range or different scan angle) are combined with all the benefits from ALS (Mandlburger et al., 2015a).
While there are numerous projects of studying geomorphic impacts with the help of multiple terrestrial and airborne LiDAR data (Avian et al., 2018(Avian et al., , 2020Eitel et al., 2016;Goodwin et al., 2017;Jones et al., 2015;Niculiț a et al., 2020;Schaffrath et al., 2015;Victoriano et al., 2018), there are only a few applications with ULS data since its development has only started over the last few years.
Both ALS and ULS systems use a laser scanner combined with a Global Navigation Satellite System (GNSS) and an inertial measurement unit (IMU). A LiDAR sensor transmits a light pulse towards the surface and measures the time the pulse takes to the surface and back. The sum of reflected pulses forms a three-dimensional (3D) point cloud. The spatial location of the individual points is determined by the position of the sensor platform, the propagation time, and the angular measurements of the laser pulse to the surface (Fujii & Fukuchi, 2005;Shan & Toth, 2017;Vosselman & Maas, 2010).
The actual measurement principle of both systems is congruent, but due to shorter sensor-object distances, the resulting point clouds of ULS provide higher point densities. Different scanning geometries (e.g., larger angle of incidence) caused by the lower/different flight height of the UAV also lead to smaller uncertainties and a more precise representation of the surface (Bakuła et al., 2017;Bakuła et al., 2020;Davidson et al., 2019;Mandlburger et al., 2015a;Pilarska et al., 2016). ULS sensor systems perform at higher accuracy, precision, laser pulse repetition rate and scan angle ranges (330 vs. 45 /60 in ALS). Compared to ALS, the processing of ULS point clouds is more complex due to higher point density, small footprints caused by the lower flight height and, resulting from this, the higher level of detail of geomorphic structures.

| Error sources and uncertainties in LiDAR data
Geomorphic impact studies and the calculation of volumetric changes need a rigorous analysis of error sources and uncertainties. This is crucial to avoid a propagation of errors and to distinguish real geomorphic changes from changes due to different errors in the data to derive a reliable change detection (Anderson, 2019;Bangen et al., 2014;Brasington et al., 2000;Cavalli et al., 2008;Lane et al., 2003;Wheaton et al., 2010).
Our work aims to identify and quantify errors and uncertainties and subsequently discuss the comparability of multi-temporal LiDAR data (point clouds and calculated DTMs) of different origin by evolving the approach of Wheaton et al. (2010). Wheaton et al. (2010) presented two robust methods on a cell-by-cell basis to estimate uncertainties in DTMs (respectively digital elevation models [DEMs]) and in a further step apply these uncertainties to geomorphic change detection by applying a lower threshold to the amount of surface change. The first approach uses a fuzzy inference system to calculate the spatial variability of uncertainties in multi-temporal DTMs; the second approach modifies this estimation by discriminating uncertainties in DTMs of difference (DoD) on the basis of the spatial coherence of erosion and deposition units. Another advanced approach to estimate uncertainty by subdividing uncorrelated, correlated, and systematic errors was proposed by Anderson (2019). However, for the catchment-wide scale of our work we decided to evolve the approach of Wheaton et al. (2010).
In our study, we compared and analysed ULS and ALS data of different point cloud densities derived with different LiDAR sensors in our study area Schöttlbach valley, a torrential catchment in Styria (Austria). Two additional evaluation areas were used to verify and improve our uncertainty analysis. The quality of LiDAR-point clouds, processed DTMs, and their uncertainties are mainly influenced by the acquisition method, the processing method of the point clouds and DTMs and the spatial surface properties of the terrain. A short summary of these error sources in LiDAR data is given in Table 1. Data acquisition errors can be divided into instrumental, methodical errors and random errors. Instrumental errors are caused by the sensor system and by the used aircraft. The laser scanning system is limited by its internal accuracy and precision (Table 2) as the inherent system bias can lead to distortions in all three dimensions. Further errors can be caused by temperature, air quality and moisture (Wise, 2000). Transmitted vibrations from the aircraft are responsible for artefacts and errors in the final point clouds. Errors in flight planning, like the adjustment of flight height, scan angle, operation time, and flight direction, strip overlapping, point density, data gaps and operator errors are summarized to methodical errors. While instrumental and random errors should not be neglected, they are only marginally addressed in this work. Random errors are random noise and have to be filtered out of the point cloud. Instrumental and methodical errors are linked to topographic complexity (Chaplot et al., 2006) and are therefore reflected in the results of our uncertainty analysis.
Errors in the processing procedure are divided into point cloud and DTM-based errors. Examples of point cloud errors are incorrect geo-referencing of the point cloud, inadequate strip adjustment, incorrect or poor filtering, random noise, and (mis-)classification of points in ground and non-ground points (Aguilar et al., 2005;Bater & Coops, 2009;Pfeifer & Mandlburger, 2017;Polat & Uysal, 2015;Sithole & Vosselman, 2004). Spatial properties of the terrain are split into surface and physical errors.
Terrain slope, roughness, curvature, low vegetation and forest cover in connection with operation time and flight height as well as the reflectivity of a surface or surface characteristics like object size, albedo, and reflectivity are the most important factors influencing the vertical and horizontal accuracy of a DTM and spatial pattern of errors (Avianet al., 2018;Chaplot et al., 2006;Hyyppä et al., 2005;Mandlburger et al., 2009Mandlburger et al., , 2015bMandlburger et al., , 2015cStere nczak et al., 2016).
Slope characteristics or micro-relief structures characterize the terrain and affect data acquisition as well as data processing. These measured points in rough terrain caused by small-scale structures can be spatially unrepresentative and lead to errors in final models. Steep slopes, however, are causing irregularities in signal reflection. Signal attenuation and fallout are forced by varying reflectivity (wavelength of the sensor system) of different objects, land-cover types and surfaces (Fisher & Tate, 2006). For example, data collected from scanners which broadcast in the near-infrared (NIR)-range may have data gaps or low ground point densities over water, snow and ice-bodies, in addition to areas with a dense vegetation cover (Lallias-Tacon et al., 2014).
Based on the detailed analysis of different error sources, a comprehensive DTM uncertainty workflow is presented in our article. We introduce a workflow for assessing and evaluating quality, uncertainties, and comparability of DTMs used for geomorphic impact studies, identifying parameters and requirements which significantly affect the quality and the comparability of DTMs. Furthermore, we discuss data acquisition and processing errors of our data pool. The main uncertainty analysis is based on data on sediment mobilization in a highly active torrential catchment (Schöttlbach creek, Styria, Austria).
The detailed outcome of this study in terms of sediment budgeting is presented in a second article in the same volume (Krenn et al., in review). The developed workflow was applied to data from two further catchments in Styria. The link between the examples presented is that in all cases, DoD were derived from datasets of different origins and densities.
The overarching research questions for our study are: • How can errors and uncertainties in calculated DTMs be assessed and quantified to serve as a basis for accurate and more reliable geomorphic impact studies?
• Which uncertainties and challenges arise from comparing ALS and ULS datasets, exemplified by our case study in the catchment area Schöttlbach?
T A B L E 1 Summary of the error sources in LiDAR (light detection and ranging) data • What are the advantages and disadvantages of using UAV-borne LIDAR point clouds in geomorphic impact studies?

| STUDY AREAS
In this article, we subdivided our study areas in a training area (Schöttlbach catchment) in which the workflow for data evaluation was developed, and two evaluation areas (Lorenzerbach and Rettenbach catchments). The detailed geomorphological analysis of the Schöttlbach data is presented in a second article (Krenn et al., in review).

| Training area
The Schöttlbach catchment is located near the small town of The Schöttlbach catchment was used as training area to develop our uncertainty workflow.

| Evaluation areas
In order to assess the performance of the workflow and its applicability for geomorphic impact studies, additional datasets from other alpine catchments with recent geomorphic activity were used  Table 2.

| MATERIALS AND METHODS
In the Schöttlbach area, the surface of the lower part of the The ALS-derived point cloud of autumn 2012 was used as reference.
Change detection analysis was carried out using the QuantumGIS (QGIS) 3.16.1 Change Detection Tool (QCD-tool) developed by the first author for the presented study (Krenn et al., in review). This open-source python-based change detection tool, which is part of the pyAlpineRisk toolbox (available via GitHub), is used to detect, calculate, and visualize surface and volume changes using two discrete DTMs. It was designed to identify errors and quantify uncertainties in multi-temporal LiDAR data of different origin, thus enabling accurate change detection and reliable volume calculations from generated DTMs as well as point clouds. For this purpose, results from our DTM uncertainty analysis described in this article can be included in the QCD-tool.

| Data acquisition
The All point clouds were collected during dry, snow-and ice-free weather conditions.
For the evaluation areas, the older ALS point clouds were also used as reference dataset ( Figure 2 and ±3 cm) and horizontal accuracy (RMSE: ± 6 cm). All point clouds were collected during dry, snow-and ice-free weather conditions.

| Data processing
The main data processing (georeferencing, strip adjustment, filtering, Based on this basic pre-processing, an additional quality check was carried out by us to prepare data for further analysis and ensure data comparability. The ULS point cloud was referenced at University of Innsbruck by real-time kinematic (RTK) positioning without using ground control points (GCP). Stable surfaces (e.g., roof areas) within the ALS point cloud were used for matching the ULS to the ALS point cloud. This means that we used the Schöttlbach-ALS as reference data for further processing (additional height adjustment) and quality control of the ULS point cloud. We used stable surfaces with different orientations and evenly distributed over the study area to verify the relative vertical and horizontal accuracy of the ULS point cloud.
For this purpose, we applied the software package LAStools for classification, filtering, height above ground adjustment of the raw point clouds and interpolation of the respective DTMs from the ground points. LAStools (rapidlasso GmbH, 2021) can process large datasets with minimal computer processing power and in a short amount of time.
Since we used LAStools for data processing, we had to use either a TIN approach with standard linear interpolation (las2dem/ blast2dem) or a grid-based approach by computing the highest, lowest or average-z-value of all ground points within a cell (lasgrid). We used the TIN approach according to Fuller and Hutchinson (2007) who used this approach for fluvial environments. The spatial resolution of the raster had to be approximated to the underlying continuous terrain as well as the point density to reduce errors in the final DoDs (Fisher & Tate, 2006). Based on the different point densities of the data and evaluating QCD-results determined with different spatial resolutions, a 0.5 m spatial resolution provided the best results for the Schöttlbach DTMs.
The software was originally designed for ALS point cloud processing; the ULS dataset required more manual post-processing (e.g., manual filtering and classification) to guarantee the best possible results. Therefore, a manual re-classification of the ULS ground points focusing on the small-scale structures along and in the riverbed was required. The quality of the DTM and in a further step any kind of hydro-morphological analysis would be substantially affected if misclassified objects were not corrected (e.g., boulders classified as vegetation). The surface roughness σ rough of landforms and that is of natural targets (geomorphic structures) is the most important terrain parameter influencing the absolute DTM height accuracy (Mandlburger et al., 2015b).

| Spatial properties of the terrain
The spatial geomorphometric properties of the terrain (mean slope, aspect, terrain roughness, elevation range, etc.) were calculated by applying a GIS-based terrain analysis (Figures 1 and 8

| Analysis of error sources and uncertainties in multi-temporal data
Due to the complex morphology in the study areas and heterogeneous quality and accuracy of the available point clouds and DTMs, a robust approach for the assessment of the comparability of multitemporal DTMs is needed. In the case of geomorphic impact studies, it is necessary to distinguish between actual surface changes and an inherent noise in the datasets in order to calculate accurate erosion and deposition volumes.
Based on the approaches of Voltz and Webster (1990), Kraus et al. (2006), and Wheaton et al. (2010), we developed a workflow for identifying and assessing errors and uncertainties in DTMs (Figure 3).
Three essential steps were taken for a robust error and uncertainty analysis in DTMs and DoD: • Detecting possible sources of error in the different datasets, • transferring these results into the DoDs, • estimating model uncertainties by calculating cell by cell deviations.
The results of the uncertainty analysis gathered during this process was transferred into the geomorphic impact study to estimate upper and lower thresholds of the sediment budget (Krenn et al., in review).
As recommended by Reuter et al. (2009) and Wasklewicz et al. (2013), we analysed all different error sources described in Section 1 (Table 1) with the help of visual analysis, evaluation of the information provided by the operating LiDAR companies and LAStools. We tried to keep errors as low as possible with regard to data processing and quantified all remaining uncertainties by using a GIS-supported statistical-empirical modelling approach on a cell-by-cell basis described in this section. This analysis was done in two main steps: Step 1 -DTM height accuracy estimation; Step 2uncertainty analysis.

| DTM height accuracy analysis
For a realistic DTM height accuracy estimation, the calculation of the accuracy in height of each interpolated raster cell is essential (Kraus et al., 2006). For our study, we assumed that the relative accuracy of the interpolated points is (1) influenced by point density and point spacing of the underlying original point cloud; (2) the absolute accuracy of the original points; (3) different terrain characteristics like steep terrain, micro-relief structures, or terrain roughness (Hyyppä et al., 2005;Kraus et al., 2006;Stere nczak et al., 2016). In terms of terrain characteristics, it is assumed that DTM height accuracy errors are higher, the more complex terrain structures are (Liu & Jezek, 1999 The classification, which was needed for the uncertainty analysis, is based on the precision/accuracy of the scanner and a detailed evaluation of the DTM height accuracy models (Table 2).
Our approach also provides accurate information on surface roughness σ rough (Bater & Coops, 2009), which also depends on the point density of the original point cloud, the underlying terrain and the spatial resolution of the DTM (Bater & Coops, 2009;Mandlburger et al., 2015b). In addition, our approach also shows how well the individual flight strips are adjusted to each other.
F I G U R E 3 Uncertainty analysis in multi-temporal data: schematic workflow of the uncertainty analysis used for geomorphic impact studies T A B L E 3 Uncertainty analysisfour input fuzzy inference system for digital terrain model (DTM) uncertainty with 51 rulesets. The four inputs are root mean square error (RMSE), point density, DTM height accuracy and data gaps (slope) (n.c., not considered)

| Uncertainty analysis
The values of uncertainty of each single raster cell within DoDs were analysed, combined, and classified making use of the following information: (1) RMSE of the discrepancies in height of the used DTMs, (2) point density, (3) slope, (4) data gaps, and (5)  with a mean ground point density of 8 pts/m 2 is ±15 cm ( Table 2; the models can therefore differ from each other by up to 30 cm) we decided to use 30 as class limit for slope.
5. Both point clouds were analysed by their point failures. Areas bigger than 6 m 2 were considered as data gaps. These were further classified into gaps in flat or moderately steep (≤ 30 ) and gaps in steep terrain (> 30 ). This classification was based on the assumption that the influence of data gaps in flat or moderately steep areas on quality and uncertainties is negligible.
6. We used the DTM height accuracy map for both DTMs described in Section 3.4.1.
The information 1-5 for each raster cell served as input parameters for the final uncertainty analysis, which was done using a further development of the fuzzy inference system described by Wheaton et al. (2010). The aim of this analysis was to assign an uncertainty class to get the maximum height error for each raster cell. This threshold can be subsequently considered in the DoD and the geomorphic change detection (QCD-tool).
Each raster cell was assessed by analysing correlations of the individual input parameters RMSE, point density, slope, data gaps, and DTM height accuracy. Therefore 51 rulesets assigned to one of the four uncertainty classes were defined and provided with a class value (Table 3). Since each raster cell can be represented by several rulesets, a fuzzy membership to the four uncertainty classes (Wheaton et al., 2010;Zadeh, 1965) was defined ( Figure 4, Table 3).
For each cell all class values were summarized and divided by the number of the applicable rulesets. Figure 4 shows the final assignment of the values to one of the four uncertainty classes by its percentage of membership.
This workflow allowed modelling of a certain degree of fuzziness and enabled a transparent assignment of each raster cell to one of the four uncertainty classes by defining its percentage of membership (Table 3).

| RESULTS
The first step for a robust error and uncertainty analysis was to detect possible sources of error in the different datasets (cf. Section 3.4). The detailed evaluation of the uncertainty sources of the data used for this study is summarized in Table 4. Further accuracy analysis on the point clouds and DTM's (relative vertical accuracy, height residuals, percentage distribution of uncertainty classes by slope) was compiled in the Appendix.
According to our detailed evaluation, most of the errors detected in the data can be attributed to methodical errors (data acquisition).
The different mean ground point densities together with different point spacings, footprints, flight heights, and the percentage of data gaps make the main differences between the compared LiDAR data and are therefore the main uncertainty sources of the data used in this study. The mean point densities of the different point clouds for example vary from 8 pts/m 2 (Schöttlbach-ALS) to 60 pts/m 2 (Schöttlbach-ULS). This is due to the different LiDAR sensors, flight and 7) resulting from the DTM interpolation, and the raster resolution must be taken into account (Table 4, (Table 4, spatial properties of the terrain).

| Uncertainties in multi-temporal data
The show that approximately 90% of all raster cells are assigned to minor uncertainties, which indicates a good comparability of the multitemporal DTMs (Figure 9). However, the riverbed, which in many studies is the main area of interest, is characterized by a slightly higher uncertainty and must therefore be considered more critically. In addition, slope areas (slope > 30 ) are also characterized by higher uncertainties ( Figure 9; Appendix Figure A3 and Table A2).

| Evaluation of uncertainty sources of evaluation areas
In the Lorenzerbach catchment, the difference of the mean ground point density between the older (8 pts/m 2 ) and the more recent sur-

| Applicability of the uncertainty analysis for geomorphic impact studies and volumetric change
For volumetric change calculations, a maximum height error for each raster cell and each uncertainty class was defined ( Table 5). The maximum height error for raster cells with minor uncertainty is based on the DTM height accuracy analysis described in Section 3.4, step 1. This was done for each study/evaluation area individually, which is the rea-  Compared with the Schöttlbach-ALS, Schöttlbach-ULS shows high point density and significantly less data gaps due to the high scan rate. However, due to the lower and variable flight height, and the different scan angle of the UAV, shading effects were caused on the side facing away from the sensor even though the data were collected in late autumn during leaf-off conditions (Figure 8d) (Zhou et al., 2020).
This results in a slightly irregular planimetric point spacing over the whole area. In principle, this problem could be avoided by adding additional flight trajectories (Brede et al., 2017;Mandlburger et al., 2015a).
Rettenbach-ALS2 was also acquired during leaf-off conditions.
The ALS flight strips were not only acquired with a high overlap of Physical errors Signal attenuation and point failures at different surfaces (e.g., water, ice, etc.) caused by the used wavelength Figure 8 50% and a strip spacing of 500 m between the parallel strips, but data were also collected by using a cross-flight strategy. When measuring a smoother terrain such as the Rettenbach catchment, a higher point density is achieved showing a regular planimetric point spacing and less data gaps. This resulted in a high-quality dataset with few data gaps and objects and structures acquired from various directions.

| Error sources and uncertainties in data processing
The results of the accuracy analysis show a high relative vertical and horizontal accuracy of the point clouds (RMSE -Schöttlbach: 0.07 m; Lorenzerbach: 0.10 m; Rettenbach: 0.07 m; Table A1). The slightly lower accuracy of Lorenzerbach-ALS2 is attributable to the boundary of the surveyed area being located along the high alpine ridges of the catchment area. In this remote terrain, less stable surfaces can be used/found for quality purposes, leading to lower accuracies (see also  (Figures 5 and 7).
We used DTMs instead of point clouds for change detection due to different point densities of the datasets. Therefore, larger uncertainties caused by DTM interpolation were accepted for a good comparability of the multi-temporal data ( Figure 10).
The high point densities of ULS point clouds is not only an advantage, but can also cause problems, for example misclassifications of small-scale height differences of rocks and boulders as non-ground points ( Figure 5) (Zhou et al., 2020). Kraus and Pfeifer (2001) also found that the combination of very dense datasets and areas with low vegetation and high surface roughness can lead to misclassifications of ground and non-ground (vegetation) points.
If used for DTM interpolation, these falsely classified ground points would cause errors in volume quantifications.
Therefore, a manual re-classification of the ULS ground points focusing on the small-scale structures along and in the riverbed is The study area and evaluation areas vary in levels of terrain roughness because of different geological properties, land-cover and land-use properties (Figures 1 and 2) These two characteristics are causing a rough terrain in the Lorenzerbach catchment. This is also reflected in the results of the DTM height accuracy and uncertainty analysis (Figures 7, 8, 9 and 10).
Therefore, higher maximum height errors for each uncertainty class were accepted for the Lorenzerbach catchment (Table 5)

| Geomorphic impact studies and volumetric change
Results of our geomorphological change detection approach underline the important impact of a detailed uncertainty analysis. By using the raw change detection, volume changes are highly overestimated.
Thresholding the studied area with a mean height error would save a lot of extra work, but degrades estimates of net change by missing/ removing some real geomorphic change ( Figure 11) (Anderson, 2019).
In our study areas the differences between thresholding versus quantifying uncertainty in change detection is substantial. A detailed analysis of error propagation and the role of thresholding in topographic change was done by Anderson (2019) Nevertheless, data processing can either reduce different errors but also cause further problems such as a deterioration of an absolute and relative accuracy or a blurring of terrain breaklines. A detailed evaluation of the original point clouds by visual but also statistical approaches (data acquisition, data processing, spatial properties of the terrain) provide helpful insights into the comparability of the used data.
The high point density of the ULS point cloud does not necessarily improve the analysis of the geomorphic impact study, because the detailed information in a ULS point cloud (e.g., micro-relief, land-cover type) makes classification of the point clouds and geomorphic interpretation challenging and makes it harder to compare the surface models with those from lower-resolution ALS data. Therefore, it is questionable whether the high level of detail in ULS datasets really adds value to geomorphic impact studies at catchment scale, as the full potential of these data can only be exploited when being compared to a dataset with similar accuracy and quality. This could also be a high-resolution, preferably cross-flight ALS dataset, as underlined by the satisfactory results of Rettenbach-ALS2.
In our change detection analysis for the three catchments, we calculated sediment relocation of 130,000 m 3 , 90,000 m 3 and 800 m 3 , respectively, using our novel uncertainty analysis workflow. All values are well in the magnitude of previous expert estimates. Using raw data or relatively low change detection thresholds (e.g., 0.1-0.2 m) results in a considerable over-estimation of sediment volumes, while a rigid higher threshold (e.g., 0.5 m) results in presumably too low volumes, as real surface changes in areas with good data quality are unnecessarily omitted. Our dynamic threshold based on uncertainty calculations could therefore be an improvement on the approach of Wheaton et al. (2010).

ACKNOWLEDGEMENTS
We kindly thank the GIS Steiermark and the Stadtvermessungsamt Graz,

CONFLICTS OF INTEREST
The authors declare that there are no conflicts of interest that could be perceived as prejudicing the impartiality of the research reported.

DATA AVAILABILITY STATEMENT
The raw data used in this study are available from third parties (GIS Steiermark, Stadtvermessungsamt Graz and WLV). The availability of these data, which were used under license for this study, is subject to restrictions.

ORCID
Nicole Kamp https://orcid.org/0000-0002-8828-6343 Paul Krenn https://orcid.org/0000-0002-2613-730X Michael Avian https://orcid.org/0000-0003-2648-9505 Oliver Sass https://orcid.org/0000-0002-9288-0724 T A B L E A 1 Relative vertical accuracy (m) of the study area and the two evaluation areas. Parameters like root mean square error (RMSE), mean error (ME), standard deviation (SD) or mean absolute error (MAE) are used to assess the precision of a dataset (Desmet, 1997;Chaplot et al., 2006;Fisher & Tate, 2006;Reuter et al., 2009Reuter et al., , 2009Wasklewicz et al., 2013). The RMSE is the most common metric used to characterize vertical errors and represents a direct comparison between measured or calculated (ZÞ to reference height values (Z ref Þ for a sample of n points (Reuter et al., 2009) Relative vertical accuracy ( . The biggest share of raster cells has a minor to low uncertainty. At Lorenzerbach a slope of > 30 lead to slightly higher uncertainties (low to acceptable). Rettenbach has a more gentle and flat terrain with raster cells with minor to low uncertainty. A good applicability of these multitemporal data for change detections in these torrential catchments is given. [Color figure can be viewed at wileyonlinelibrary.com]