HIGH-RESOLUTION DEBRIS FLOW VOLUME MAPPING WITH UNMANNED AERIAL SYSTEMS (UAS) AND PHOTOGRAMMETRIC TECHNIQUES

Debris flows cause an average € 30 million damages and 1-2 fatalities every year in Austria. Detailed documentation of their extent and magnitude is essential for understanding, preventing and mitigating these natural hazard events. The recent development of unmanned aerial systems (UAS) has provided a new possibility for on-demand high-resolution monitoring and mapping. Here, we present a study, where the spatial extent and volume of a large debris flow event were mapped with different UAS, fitted with commercial off-the-shelf sensors. Orthophotos and digital terrain models (DTM) were calculated using structure-from-motion photogrammetry software. Terrain height differences caused by the debris flow in the catchment and valley floor were derived by subtracting the pre-event airborne laser scanning (ALS) DTM from a post-event UAS-DTM. The analysis of the volumetric sediment budget showed, that approximately 265,000 m3 material was mobilised in the catchment, of which 45,000 m3 settled there; of the material, which reached the valley floor, 120,000 m3 was deposited, while another 10,000 m3 was eroded from there. The UASresults were validated against ALS data and imagery from a traditional manned-aircraft photogrammetry campaign. In conclusion, the UAS-data can reach an accuracy and precision comparable to manned aircraft data, but with the added benefits of higher flexibility, easier repeatability, less operational constraints and higher spatial resolution. * Corresponding author


INTRODUCTION
Debris flows are described as 'rapid, gravity-induced mass movements consisting of a mixture of water, sediment, wood and anthropogenic debris that propagate along channels incised on mountain slopes and onto debris fans' (Gregoretti et al., 2016).These natural hazard events have a substantial impact on the quality of life in mountainous regions: In Austria, the examination of 5,000 debris flows, recorded between 1972 and 2004, resulted in a total estimated damage of € 965 million and 49 fatalities (Oberndorfer et al., 2007).Promptly mapping the consequences of these events by determining the spatial extent and volume of eroded and deposited material, is highly relevant to scientists and practitioners, e.g. for numerical simulation modelling (Rickenmann et al., 2006;Han et al., 2015), process understanding (Theule et al., 2015;Pellegrino et al., 2015), and natural hazard management (Ballesteros Cánovas et al., 2016;Aronica et al., 2012).Conventional mapping techniques mostly require personnel to directly access the process area (catchment and valley floor), which is a hazardous or even impossible task.Additional data acquisition from terrestrial (e.g.terrestrial laser scanning), airborne (e.g.traditional photogrammetry from manned aircraft) or spaceborne sources (e.g.high-resolution satellite imagery) is typically only commissioned in case of large-scale events (Linder et al., 2015).
In recent years, the development of unmanned aerial systems (UAS) (Colomina & Molina, 2014) has provided a wide range of new possibilities for high resolution monitoring and mapping (Lucieer et al., 2014).UAS are able to bridge the gap between full-scale, manned aerial, and terrestrial observations (Briese et al., 2013;Rosnell & Honkavaara, 2012).They are credited as being able to supply on-demand imagery at an unprecedented level of detail (ground resolution of few centimetres or millimetres) (Lucieer et al., 2014), in a cost-efficient and flexible manner, albeit over small areas (<1 km²) (Bühler et al., 2016;Ryan et al., 2015).Additionally, the development of novel computer vision techniques (structure-from-motion) and their implementation into commercially available software packages (e.g.Agisoft Photoscan Pro) have reduced the requirements for the recorded data (Vander Jagt et al., 2015;Turner et al., 2012).This development makes high-performance algorithms for 3D-processing more readily available to the UAS-community (Mancini et al., 2013;Hugenholtz et al., 2013).
The objective of this paper is to present the use of UAS-data to map the extent and calculate the volumetric sediment budget of a debris flow event, i.e. the total volume of eroded and deposited material in the catchment and valley floor.

STUDY SITE
The study site is situated in the Sellrain Valley of the Stubaier Alps, central Tyrol, Austria (47°12' N, 11°12' E) (Figure 1, top left).It is a tributary to the Inn Valley and has a population of approximately 2,200.The valley's V-shape narrows down the space available for settlements and infrastructure, which are often endangered by natural hazards (Bitschnau & Obermair, 2012).The debris flow originated in the catchment of the Seigesbach torrent, a tributary to the Melach -the main river draining the valley.Prior to 2015, several smaller events and one with comparable magnitude (in 1928) had been recorded (Fliri, 1998).Therefore, no mitigation measures had been constructed in the Seigesbach torrent.The entire catchment comprises an area of about 4 km² and reaches its highest point at the Fotscher Windegg (2,577 m a.s.l.).The land cover in the catchment is dominated by spruces (Picea abies) and alpine meadows.The majority of the debris was deposited in the valley floor in the western part of the village Sellrain.The valley floor lies at approximately 950 m a.s.l. and is mainly characterised by meadows, with small clusters of buildings and patches of wooded areas.The main road connection through the valley (Sellrainstraße, L13) runs along the upper section of Figure 1, top right.The areas of interest (AOI), where the UAS-data was collected and the centre of Sellrain are also indicated in this figure.

Debris Flow Event Description
In May 2015, the precipitation rate in the study area exceeded the long-term average by approximately 80% (Drechsler, 2015).During the last days of May, almost the entire catchment was covered by snow.By 6 June 2015, one day prior to the event, the entire snow cover had melted.This snowmelt, combined with elevated precipitation leading up to the event, resulted in heavily waterlogged ground in the catchment.During the evening of 6 June, extraordinarily high amounts of precipitation reached the catchment.On 7 June, an intense, almost stationary thunderstorm with hail occurred, triggering the debris flow.The precipitation reached up to 110 mm in six hours (Drechsler, 2015).The event culminated in a large debris flow, covering parts of the valley floor with a big alluvial cone in a comparatively short time.Two buildings were totally destroyed and a further 15 damaged; three road bridges were damaged or destroyed, the main road connection was impacted over 300 m length and interrupted for several months; four hectares of green and willow land was affected; no damage to persons occurred, but livestock was lost (Jenner, 2015).The authorities later estimated a total damage of approximately € 30 million (ORF, 2015).On all flights, the cameras' shutter speed, aperture and ISO were chosen to maximise image quality and avoid motion blur (1/400-1/640, f/4-f/18, ISO 400-1000, respectively).Table 1 summarises the details of the six UAS-flights at the study site.
The UAS both featured integrated global navigation satellite system / inertial measurement unit (GNSS/IMU) components, connected to an autopilot, and were therefore capable of automatically following a pre-defined flight path.The technical specifications of the UAS and the deployed sensors are provided in the Appendix - As the UAS were not equipped with high-quality (i.e.surveygrade) GNSS receivers to determine their location in space, indirect georeferencing with ground control had to be used (e.g.Nolan et al., 2015;Harwin et al., 2015).Prior to the UASflights, 30 ground control points (GCP), consisting of 0.4 x 0.4 m black and white checkered wooden boards were placed at predefined locations in the catchment and valley floor AOIs, which were not affected by the debris flow.The position of the GCPs was recorded using a terrestrial GNSS (Trimble GeoExplorer XT 2008, expected accuracy x/y-direction 0.1, zdirection 0.2 m) with an external antenna (model: Hurricane), capable of receiving single-frequency GPS (global positioning system) data.Real-time correction was used in the field and differential correction from the scripps orbit and permanent array centre in Bolzano (80 km from the study site) applied in the office via the Trimble software Pathfinder Office (Trimble, 2015).The final x-, y-& z-coordinates of each GCP were averaged from the >200 points recorded at each GCP location.
Data on the height of the pre-event terrain was available from two ALS campaigns commissioned by the Province of Tyrol: the catchment was mapped in 2009 with a ground sampling distance (GSD) of 1 m, the valley floor in 2014 with a 0.5 m GSD.

Photogrammetric reconstruction
The UAS-images were processed with Agisoft's Photoscan Pro (version 1.1.6)(Agisoft LLC, 2016), a commercially available photogrammetric software suite, that is widely spread in the UAS-community (Tonkin et al., 2014).It is based on a structure-from-motion algorithm and provides a complete, photogrammetric workflow, with particular emphasis on multiview stereopsis (Harwin et al., 2015): i) tie point matching; ii) bundle adjustment (here constrained by assigning high weights to the GCP coordinatesknown as indirect georeferencing or conventional aerotriangulation (Vander Jagt et al., 2015)); iii) linear 7-parameter conversion; removal of non-linear deformations; iv) dense point cloud (DPC) generation with multiview stereo reconstruction; v) export of georeferenced orthophotos and digital elevation models (DEM).We referenced all data to the respective standard national coordinate systems (EPSG-Code 31254, Gebrauchshöhen Adria).
The height values of the DEMs derived from the photogrammetrically calculated DPC (Figure 3, left) generally refer to the height above sea level of the terrain, buildings or vegetation, captured in the scene.It is therefore more accurate refer to them as digital surface models (DSM) (Weibel & Heller, 1991).When estimating the deposition and erosion volumes of the debris flow, we calculated the change in terrain height.We therefore removed all objects above or on the terrain, from the DSM, resulting in a digital terrain model (DTM) (Weibel & Heller, 1991).The DTMs were generated by classifying the DPC in Photoscan into ground and non-ground points (Figure 3, right).For those areas, where no ground points are available, Photoscan performs a linear interpolation of the surface.
The resulting orthophotos (Figure 4, left) and DSM (Figure 4, right) of the valley floor and catchment, featured a GSD of 0.02/0.05m and 0.08/0.2m, respectively, with lower average GSD in the catchment.The mean effective overlap of the UASdata was 8-16 images.The DEMs were calculated from photogrammetric DPCs, which have an average point density of 30-160 points/m².

Volume Calculation
The terrain height change caused by the debris flow was calculated by subtracting the pre-event ALS-DTM from the post-event UAS-DTM.These difference DTMs (dDTM) thus provide information on the amount of deposition (positive values) and erosion (negative values) of debris for each pixel (Nolan et al., 2015).We estimated the total volume of debris deposition and erosion, by summing up the dDTM pixel values within the AOIs.To avoid possible artefacts resulting from the DTM-interpolation, we masked out buildings and vegetation areas.To counter systematic error in the dDTM resulting from different height reference systems, we extracted the z-values of all GCPs from the ALS DTM.The total debris flow mass balance calculation showed that 265,000 m³ (±42,000 m³) material was mobilised in the catchment, of which 45,000 m³ (±13,000 m³) settled there; of the material, which reached the valley floor, 120,000 m³ (±5,000 m³) was deposited there, while another 10,000 m³ (±2,000 m³) was eroded.

Validation of Results
Accuracy and precision of the UAS-results was calculated by comparison with manned aircraft imagery, ALS data, extensive fieldwork and a second UAS-flight in the valley floor.The accuracy of the UAS-data lies within 0.05-0.15m (root mean square error -RMSE) and 0.3-0.5 m in xyz-directions, for the valley floor and catchment, respectively.The uncertainty range of the mass balance calculation above corresponds to this RMSE.The precision in the valley floor was derived from the second UAS-flightit lies at 0.05 m.A more advanced analysis of accuracy and precision is currently in preparation.

CONCLUSIONS
In this study, aerial imagery from a fixed-wing and a multicopter UAS were used to map an area of 2.8 km² with the aim of determining the volumetric sediment budget of a debris flow event in the Austrian Alps in summer 2015.Using structure-from-motion photogrammetry and indirect georeferencing, high-resolution post-event orthophotos (<0.05 m GSD) and DTMs (<0.2 m GSD) of the valley floor and catchment were generated.A calculation of the sediment budget with pre-event ALS-data could be successfully performed at an accuracy of 0.05 -0.5 m and precision of 0.05 m RMSE.The UAS-data can therefore reach an accuracy and precision comparable to manned aircraft data, but with the added benefits of higher flexibility, easier repeatability, less operational constraints and higher spatial resolution.However, the study also showed, that especially extensive (>1 km²), densely vegetated areas with large elevation differences, pose a challenge for UAS-based volume estimations, as indicated by the higher error margins in the catchment.
This study represents one of the very few published attempts to use UAS-based imagery and photogrammetric techniques to map the consequences of a debris flow event.The results from this UAS-campaign provided the authorities with a highresolution documentation of the debris flow's spatial extent and magnitude, as well as decision support for mitigation measure planning (e.g.planning mitigation measures).

Figure 1 .
Figure 1.Location of study site in Austria (top left); overview of AOI of the UAS-flights (top right); terrestrial photographs of the catchment on 18 June 2015 (bottom left) and the valley floor on 9 June 2015 (bottom right) (sources: administrative boundaries -GADM, 2016; imagery top row -ESA Sentinel 2A, 2016; photographs bottom row -Pittracher WLV, 2015)

Figure 5 (
Figure 5 (top) shows the results of the volumetric sediment budget calculation for the valley floor, coloured by erosion (blue) and deposition (brown).The majority of debris collected in the riverbed of the Melach (s-shaped feature in upper part of the figure).Here deposition depths reached up to 10 m, as also evidenced by the cross-section (Figure 5, centre left).While overall the deposition dominates, isolated patches of erosion were also detected.The highest erosion values (>4 m) are reached on the left bank of the Seigesbach torrent, which enters the figure from the bottom left.The results from the valley floor (Figure 5, bottom) highlight the extent of of material from the catchment.Large sections of this part of the catchment

Figure 5 .
Figure 5. Results of the volumetric sediment budget calculation in the valley floor (top) and catchment (bottom); two cross-sections through the pre-(ALS) and post-event (UAS) DTM (centre leftdeposition; centre rightcatchment), the location of the crosssections are indicated by the white lines in both figures -A & B / C & D marking the starting and end points, respectively

Table 1 .
Details of UAS-flights

Table 4 .
Technical specifications of the UAS-sensors