Brief Communication: An Autonomous UAV for Catchment-Wide Monitoring of a Debris Flow Torrent

. Debris flows threaten communities in mountain regions worldwide. Combining modern photogrammetric processing with autonomous unoccupied (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) unmanned (cid:58) aerial vehicle (UAV) flights at sub-weekly intervals allows mapping of sediment dynamics in a debris-flow (cid:58)(cid:58)(cid:58)(cid:58)(cid:58) debris (cid:58)(cid:58)(cid:58)(cid:58) flow (cid:58) catchment. This provides important information for sediment disposition that pre-conditions the catchment for debris flow occurrence. At the Illgraben debris-flow catchment in Switzerland, our autonomous UAV launched 5 nearly 50 times in the snow-free periods in 2019-2021 with typical flight intervals of 2-4 days, producing 350-400 images every flight. The observed terrain changes (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) terrain-changes (cid:58) resulting from debris flows exhibit preferred locations of erosion and deposition, including memory effects as previously deposited material is preferentially removed during subsequent debris flows. Such data are critical for the validation of geomorphological process models. Given the remote terrain, the mapped short-term erosion and deposition structures are difficult to obtain with conventional measurements. The proposed method thus 10 fills an observational gap, which ground-based monitoring and satellite-based (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58) satellite

flow hazards. In this way, water-damming deposits from previous debris flows or landslides can be identified, whose breaching may be particularly difficult to predict since it is not related to meteorological parameters (Godt and Coe, 2007). Similarly, repeated digital elevation models (DEMs) can reveal temporal exhaustion of sediments available in debris flow source areas. 20 This "supply limitation" temporarily lowers the debris flow hazard, in contrast to sudden slope failures whose deposits in torrent beds suddenly increase the hazard (Bovis and Jakob, 1999). Such variations in sediment availability may explain why rainfall thresholds tend to perform poorly in terms of warning (Cannon et al., 2008;Kean et al., 2012;Gregoretti et al., 2016;Rengers et al., 2016). However, the spatial coverage and the temporal resolution (typically on the order of tens of km 2 and days to weeks, respectively) needed to reliably monitor an entire catchment requires costly surveillance flights or time demanding 25 site visits.
Here we introduce a new approach for monitoring sediment changes in catchments that are prone to debris flows. Using an autonomous unmanned aerial vehicle (UAV) performing flights at intervals as low as a few days, we generate time series of DEM differences for a Swiss torrent. We employ a recently developed photogrammetric processing scheme to identify terrain changes in the hillslope-channel area with decimeter precision, showing erosion and deposition patterns related to :::: both :::::: during 30 debris flows and to :::: from lateral slope failures. We propose to integrate the system into multisensing monitoring approaches to optimize the assessment of debris flow hazards in otherwise difficult-to-access mountainous regions.

Study Site: Illgraben, Switzerland
This study focuses on the Illgraben torrent in Switzerland's Canton Valais (VS), which drains a 9 km 2 catchment ( Figure   1) and produces around 5 debris flows per year on average reaching the channel outlet at the Rhône River (Badoux et al.,35 2009). With little sediment discharge outside debris flow episodes, Illgraben delivers annually some 10 5 m 3 of material to the Rhône (Hirschberg et al., 2021a). Mobilized into debris flows during intense summer precipitation, sediment deposits within the upper channel sections are supplied from frost-weathering slope failures (Bennett et al., 2013;Hirschberg et al., 2021b, a).
Probabilistic modelling indicates that sediment supply limitations affect the formation of large debris flows, although there are no direct observations to confirm this (Bennett et al., 2014;Hirschberg et al., 2021a).

40
Entrainment and positive feedback between sediment motion in the channel and on lateral slopes result in debris flow volumes that may exceed 10 5 m 3 and thus the volumes of individual rockfalls and landslides feeding the channel (Schlunegger et al., 2009;Berger et al., 2011;Burtin et al., 2014). As an exception, in 1961 a rock avalanche filled the upper channel section with 3.5 × 10 6 m 3 of sediments (Gabus et al., 2008  . As a result, an alarm system ::: An ::::: alarm :::::: system :::: thus : signals in-torrent detection of mass movements at the Illgraben mouth (Figure 1; Badoux et al., 2009). Additional instrumentation near the channel outlet includes flow depth gauges, a force plate for instantaneous flow weight measurements, automatic cameras, infrasound microphones, and seismometers (e.g., McArdell et al., 2007;Marchetti et al., 2019;Schimmel et al., 2018;Chmiel et al., 2021) :::::::::::::::::::::::::::::::::::: (McArdell et al., 2007;Marchetti et al., 201 . Repeated topographic surveys generating DEMs have been used to study the controls on erosion and deposition by debris flows at Illgraben: a terrestrial laser scanner has been used over a 300 m long reach of the main Illgraben channel for 14 events 55 occuring between 2007 and 2009 (Schürch et al., 2011) and a UAV over a 3 km long channel section mainly on the fan, which was flown before and after six debris flows in 2018 and 2019 (de Haas et al., 2020). These studies provided highly resolved topographic data between individual debris flows and provided insights into the roles of channel geometry, check dams and debris-flow characteristics in erosion and deposition processes. For example, debris flows tend to erode at locations where the previous event was depositional, and to deposit where previous events were erosional (de Haas et al., 2020). To study 60 variability in sediment production, four aerial images recorded over 2008 and 2009 were sufficient to identify a downslopedirected sediment cascade at the seasonal scale (Berger et al., 2011). Sediment dynamics were also studied over decadal time scales (42 years in total) but at a coarser temporal resolution (6-20 years). Aerial images showed, for example, an increase in the Illgraben erosion rate from the 1980s, likely due to decreased snow cover and enhanced weathering (Bennett et al., 2013).
While these studies were helpful in describing patterns of sediment supply from hillslopes and its relation to sediment yield, 65 the mass movement initiation mechanism remains difficult to identify. Similarly difficult is the assessment of sediment budgets at the event scale, since some eroded areas may be masked in the aerial images and since the reconstructions from older images are affected by uncertainties of up to 5 m (Bennett et al., 2013).

Autonomous UAV
To monitor sediment dynamics in the upper catchment, where access on foot and in-torrent instrumentation is limited, we 70 deployed an autonomous UAV near the Illgraben mouth for several months during summers 2019-2021 ( Figure 1, Table 1).
The system was developed by the company "Meteomatics" (www.meteomatics.com) and consists of ( to a 220 V power supply and recharged the Meteodrone automatically through a specific charging port. The latter automatically connects to the UAV from a hole in the landing platform, which is sealed prior to and after charging. One fully charged battery yields about 30 min of flight time. Batteries were charged with a current of 20 A at 24 Vwhose power was provided to the Meteobase via a continuous external line power supply. A power generator has been employed in previous installations of the Meteobase. A combination of solar-panels and batteries would also be feasible but has yet to be implemented. The external 85 line power supply also . :::: This :::: also feeds the other electric consumers, such as the air conditioning unit used for climatizing the Meteobase's interior during very hot or cold days. The Meteobase also acts as an operational relay between Meteodrone and operator, which remotely supervises the flights as demanded by regulations. The Meteobase also ensures that procedures such as charging, data upload and download, UAV positioning on the landing platform, or climatizing, are performed automatically.
Meteodrone and Meteobase communicate through radio connection, whereas the base and the remote operator communicate 90 via a combination of 4G GSM and Local Area Network internet connections.
As payload, the UAV carried a Yuneec E90, 20 megapixel, nadir-oriented photocamera with an electronic shutter and a 1inch Complementary Metal Oxide Semiconductor sensor. The camera view included mainly the torrent channel and therefore lateral slopes were only surveyed near the channel. Pictures were taken every 2 seconds leading to an overlap of 70 to 80 %.
Synchronization of RTK geolocation and camera shutter was not implemented for technical reasons. This synchronization 95 is planned in future deployments. Other technical challenges included limited GNSS and GSM reception, as well as limited durability of components which required replacement. These were parts that were susceptible to environmental conditions and which needed adaptions or a replacement for more robust versions. One example is the compass, which failed at the beginning of the campaign due to direct sunlight radiation and developing heat and was replaced with a more reliable version.
Fixing such simple yet critical technical glitches was the reason why more maintenance was necessary in the beginning of the 100 project, compared to the end of the project when inspections were carried out about once every two months. In general, the Meteodrone requires maintenance on the motors after 150 flight hours, exchange of the battery after 150 recharging cycles and parachute replacement after 12 months. In total, the setup :::: This allowed for 41 autonomous flights during July-October 2020 and July-August 2021 after a test period in 2019 (Table 1).
In Interventions of the operator were limited to Abort Missions, e.g. due to strong wind, or changing weather conditions. The operator has different options to intervene, depending on the situation. In case the UAV is still capable of flight, the operator can abort the mission or return to launch. The operator was also allowed to hold the flight or descent to a safe altitude. The parachute rescue system is either launched automatically upon recognition of a problem by the onboard systems or at any time by the operator. This means that also during the starting and landing phase the operator can "kill" the system should it be 115 necessary to prevent harm. The operator can observe the landing and starting site through the surveillance camera, ensuring 1 https://www.bazl.admin.ch/bazl/en/home/good-to-know/drohnen/wichtigsten-regeln/bewilligungen/sora.html, last accessed 05.04.2022 nobody is in the immediate vicinity. We decided against a fenced base station for this project, but this may be appropriate in the future to protect the Meteobase and Meteodrone from vandalism.
Catchment-wide flights refer to the along-channel section between the Meteobase and the head of the Illgraben channel. This extent was covered by 6 km-long round-trip flights taking approximately 20 minutes each. The UAV flew at 100 m altitude 120 above the torrent channel with a speed of 5-7 m/s, taking between 350 and 400 photographs along its way.

Photogrammetric Processing& Results
Only 2 usable Ground Control Points were collected in the accessible section of the channel, 500-1000 m upstream of the UAV base. This means that the GNSS coordinates acquired by the UAV were the only reliable georeferencing information.

125
Initially, the autonomous imagery was processed by using the built-in functionalities of the software Agisoft Metashape version 1.7.0. Images were aligned for each survey individually, using their full resolution, rolling shutter compensation, an image location accuracy of 20 cm, and otherwise default parameters. However, the accuracy of the RTK GNSS positions proved to be insufficient: tie-point residual errors never converged to sub-pixel levels, which indicates a faulty camera model, most likely resulting from poor georeferencing (James et al., 2017). We attribute this inaccuracy to the insufficient synchronization 130 between the internal clock of the camera and the clock of the GNSS receiver, leading to incorrect matches with the GNSS track (c.f. Girod et al., 2017). To lower the georeferencing and tie point residual errors to reasonable levels, a more advanced alignment technique was therefore needed.
We opted for the "co-alignment" approach, proposed by Cook and Dietze (2019) for processing surveys for change detection.
In this workflow, the images from two or more surveys are pooled during Agisoft Metashape's image alignment and model 135 optimization processing steps. After the model geometry is set, the dense clouds are then calculated separately for each survey.
The identification of common tie points in stable areas visible in photographs from different surveys results in a model that fixes the different surveys with a common geometry. While this approach does not improve the accuracy of global georeferencing, the high comparative precision between co-aligned surveys makes the approach effective for constraining geomorphic change (Hendrickx et al., 2020;Watson et al., 2020). Autonomous surveys are particularly suited for this type of co-alignment 140 approach, as the photographs show a high consistency in orientation, altitude, and location from survey to survey. For our imagery, calculated elevation differences of stable terrain indicated a height error of 0.2 m for the co-alignment processing.
The processing of one flight took about four hours on a 12-core 3 GHz Intel Zeon central processing unit without graphics processing unit acceleration. which reached the channel :::::::: catchment outlet. At the foot of lateral tributary gullies, upstream erosion during winter months left deposits within the channel, with thicknesses of more than 2 m (Section A in Figure 2 : 1). Subsequent debris flows preferentially eroded these deposits, which is a manifestation of a "memory effect" that had previously been observed behind Illgraben's check dams on the debris fan (de Haas et al., 2020). In the upper catchment part surveyed in the present study, check dams are also subject to this memory effect (Section B in Figure 2 : 1). Also apparent are typical lateral levee deposits (e.g. Johnson et al.,150 2012), and erosion within the channel flow center (Section C in Figure 2 : 1).
This investigation has shown that quantitative information on catchment-wide sediment dynamics can be obtained on timescales of hours, i.e. on timescales that are only constrained by UAV battery charge and flight times. This fills a critical 160 gap left by costly airborne sensing and satellite-based methods, which have multi-day return periods.
Although the main channel was surveyed far upstream, debris-flow initiation areas and the triggering mechanism could not be identified unambiguously. It is clear, however, that lateral tributaries deliver significant amounts of sediments to the main channel, where they are temporarily stored. The re-mobilization of these deposits may contribute to debris-flow formation and subsequent supply-limitations (Berger et al., 2011). Currently, our flight planning and the nadir-looking camera view do not 165 allow quantification of the individual tributary contributions to debris flow volumes and dynamics. However, it is possible to approximately date relevant tributary activity and study its controls, which will be subject of future systematic researchusing the entire present data set. We also envision the ::: For :::::: future ::::::: research, ::: we :::::::: envision application of our method for constraining geomorphological models that describe sediment movement in response to short-term meteorological forcing.
Our application of an autonomous UAV leverages recent success of photogrammetry for monitoring geomorphological 170 processes. UAV-based photogrammetry combined with the structure-from-motion algorithm (Smith et al., 2016), which was also used in our co-alignment approach (Cook and Dietze, 2019), offers the unique advantage of mapping extended portions of changing terrain. This is a valuable alternative to DEM differences derived from terrestrial laser scanning (TLS), which has also  21.08.2020 -03.05.2021 03.05.2021 -18.06.2021 03.05.2021 -18.06.2021 18.06.2021 -23.06.2021 13.08.2020 -17.08  . Boxes A and B illustrate memory effects, with erosion concentrated in areas of previous deposition (white arrows). A and B show the cases for tributary deposition within the channel and debris flow deposition behind a check dam, respectively. Box C highlights lateral levee formation and in-channel erosion (black arrows).
Color map gives erosion (-) and deposition (+) in meters.Note that scaling factors are relative among zoom boxes. In Boxes B1, B2 and C, thin blue arrows indicate the location of check dams. been used in debris flow catchments to study erosion patterns in relation to debris flow dynamics (Dietrich and Krautblatter, 2019) , process domains of debris flow initiation (Staley et al., 2014) and channel response to climate signals (Bonneau et al., 2022) 175 . Whereas topography derived from UAV-based photogrammetry and TLS agrees within a few decimeters (Cook, 2017), the former method has the advantage of mapping poorly accessible terrain where TLS equipment cannot be installed.
Despite its advantages, UAV-based photogrammetry has difficulty in resolving terrain changes for certain types of surfaces, in particular vegetation cover (Cook, 2017). This disadvantage, however, may be negligible for active debris flow catchments, whose sediment dynamics allow for little or no sustained vegetation growth. Moreover, our application of co-alignment reduces 180 the dependence on ground control points which are also difficult to obtain in poorly accessible terrain. Consequently, the main constraints of our proposed sediment dynamics monitoring with an autonomous UAV are the dependence on a reliable power and internet connection for the UAV base as well as GNSS signal and radio connection between the base and the UAV.
In the future, autonomous UAV ::: On ::: the ::::::: technical :::: side, ::::::::::: autonomous :::: UAV operation can be linked to other sensor systems: The UAV could be sent to map runout and damage immediately after an event, which could be detected by seismic or infrasound 185 sensors, for example. These latter methods have the advantage of a large radius of sensitivity. However, accurate event location and volume estimates like UAV-derived DEM differences provide, are often unavailable for seismic and infrasound monitoring.
In the spirit of an "Internet of Things (IoT)" approach, the UAV system could be integrated within autonomous multi-sensing platforms that leverage the strengths of individual sensor components. The aftermath of the 2017 rock-avalanche at Piz Cengalo, Switzerland, underlined the urgent need for such post-event monitoring: within 1-2 weeks, unstable rock avalanche deposits 190 subject to high pore pressures produced 15 debris flows destroying parts of the village of Bondo (Walter et al., 2020). Such rapid secondary effects of the rock avalanche were not expected but in the future could be monitored and warned against with a quickly deployable autonomous UAV.
Rapid technological developments and increasing sensor coverage targeting rapid mass movements are currently preparing the ground for autonomous monitoring and warning systems for Alpine hazards. For our specific case, BVLOS flight permis-195 sions still required a human to follow the UAV operation from a remote location. Apart from legal constraints, this type of surveillance was not necessary from a technical and operational point of view. We thus anticipate that it is only a question of time until the presented technology will find its way into IoT monitoring solutions for natural hazards in Alpine terrain.
Data availability. UAV images are archived a WSL and access can be granted by the authors.
Author contributions. FW, NA and DF planned the autonomous UAV operation at Illgraben. EH, EM, KC and MD tested conventional 200 and co-alignment processing of the photogrammetric data. LE, FH and EH applied the co-alignment processing to all available images and together with FW, BM, JH and PM interpreted the results in terms of torrential activity. MF and LH lead the UAV deployment and operation.
All authors participated in the manuscript writing.