Timely prediction potential of landslide early warning systems with multispectral remote sensing: a conceptual approach tested in the Sattelkar, Austria

While optical remote sensing has demonstrated its capabilities for landslide detection and monitoring, spatial and temporal demands for landslide early warning systems (LEWS) were not met until recently. We introduce a novel conceptual 10 approach to structure and quantitatively assess lead time for LEWS. We analysed “time to warning” as a sequence; (i) time to collect, (ii) to process and (iii) to evaluate relevant optical data. The difference between “time to warning” and “forecasting window” (i.e. time from hazard becoming predictable until event) is the lead time for reactive measures. We tested digital image correlation (DIC) of best–suited spatiotemporal techniques, i.e. 3 m resolution PlanetScope daily imagery, and 0.16 m resolution UAS derived orthophotos to reveal fast ground displacement and acceleration of a deep– 15 seated, complex alpine mass movement leading to massive debris flow events. The time to warning for UAS and PlanetScope totals 31h/21h and is comprised of (i) time to collect 12/14h, (ii) process 17/5h and (iii) evaluate 2/2h, which is well below the forecasting window for recent benchmarks and facilitates lead time for reactive measures. We show optical remote sensing data can support LEWS with a sufficiently fast processing time, demonstrating the feasibility of optical sensors for LEWS. 20


Introduction
Landslides are a major natural hazard leading to human casualties and socio-economic impacts, mainly by causing infrastructure damage (Dikau et al., 1996;Hilker et al., 2009). They are often triggered by earthquakes, intense short-period or prolonged precipitation, and human activities (Hungr et al., 2014;Froude and Petley, 2018). In a systematic review Gariano and Guzzetti (2016) report that 80 % of the papers examined show causal relationships between landslides and 25 climate change. The ongoing warming of the climate (IPCC, 2014) is likely to decrease slope stability and increase landslide activity (Huggel et al., 2012;Seneviratne et al., 2012), which indicates a vital need to improve the ability to detect, monitor and issue early warnings of landslides and thus to reduce and mitigate landslide risk.
Early warning, refers to a set of capacities for the timely and effective provision of warning information through institutions, such that individuals, communities and organisations exposed to a hazard are able to take action with sufficient time to 30 reduce or avoid risk and prepare an effective response (UNISDR, 2009). According to UNISDR (2006), an effective early warning system consists of four elements: (1) risk knowledge, the systematic data collection and risk assessment; (2) the monitoring and warning service; (3) the dissemination and communication of risk as well as early warnings; and (4) the response capabilities on local and national levels. Lead time as defined in the context of LEWS is the interval between the issue of a warning (i.e. dissemination) and the forecasted landslide onset (Pecoraro et al. 2019) and thus crucially depends on 35 time requirements in phases (1)-(3). The success of an EWS therefore requires measurable pre-failure motion (or slow slope displacement) to allow for sufficient lead time for decisions on reactions and counter measures (Grasso, 2014;Hungr et al., 2014).
While remote sensing has been established for early warnings, remote sensing is not yet used for real early warnings of the onset of landslides in steep-alpine terrain (with a few exceptions), where geotechnical instruments are still preferred. 40 Exceptions include terrestrial InSAR (Pesci et al., 2011;Walter et al. 2020) and terrestrial laser scanning with high repetition rates. However, repeated UAS (unmanned aerial systems) and optical satellite images (PlanetScope) with high repetition rates have so far not been applied for landslide early warning in steep-alpine catchments. In this regard, knowledge of sensor capabilities and limitations is essential, as it determines which rates and magnitudes of pre-failure motion can potentially be identified (Desrues et al., 2019). Our proposed framework refers to mass movements in steep-alpine catchments with 45 significant pre-failure motion over sufficient time periods and thus excludes instantaneous events triggered by processes such as heavy rainfalls or earthquakes.
This study presents a new concept to systematically evaluate remote sensing techniques to estimate and increase lead time for landslide early warnings in these catchments. We do not start from the perspective of available data; instead, we define necessary time constraints to successfully employ remote-sensing data to provide early warnings. This approach reduces to a 50 small number the suitable remote sensing products with high temporal and spatial resolution. With these constraints, we investigated the application of data from satellites and UAS to allow the assessment of the data, after a spaceborne area-wide but low-resolution acquisition, into a downscaled detailed image recording. In so doing, we analysed the capability of these different passive remote sensing systems focusing on spatiotemporal capabilities for ground motion detection and landslide evolution to provide early warnings. 55 Recently, the spatial and temporal resolution of optical satellite imagery has significantly improved (Scaioni et al., 2014) and has allowed substantial advances in the definition of displacement rates and acceleration thresholds to approach requirements for early warning purposes. This is essential since spatial and temporal resolution determine whether landslide monitoring is possible with the detection of displacement rates and approximate acceleration thresholds, both of which are lacking if 60 information is based solely on post-event studies (Reid et al., 2008;Calvello, 2017). Landslide monitoring offers the potential to significantly advance landslide early warning systems (LEWS) (Chae et al., 2017;Crosta et al., 2017).
Previously, high spatial resolution satellite data was obtained at the expense of a reduction in the revisit rates (Aubrecht et al., 2017). Consequently, the return period between two images increased, limiting ground displacement assessment and the range of observable motion rates. The number of useful images was further reduced due to natural factors such as snow 65 cover, cloud cover and cloud shadows. High-resolution remote sensing data was long restricted due to high costs and data volume (Goodchild, 2011;Westoby et al., 2012). Today commercial very high resolution (VHR) optical satellites exist, but tasked acquisitions make them inflexible and very cost intensive, thus limiting research (Butler, 2014;Lucieer et al., 2014).
There is a vast spectrum of available remote sensing data with high spatiotemporal resolution (Table 1). Complementary use of different remote sensing sources can significantly improve landslide assessment as demonstrated by Stumpf et al. (2018) 70 and Bontemps et al. (2018), who draw on archive data and utilise different sensor combinations to analyse the evolution of ground motion. and a new generation of micro cube satellites, sent into orbit in large numbers by PlanetLabs Inc. These micro cube satellites, known as 'Doves'/PlanetScope (from now on referred to as PlanetScope satellites), and Sentinel-2 a/b offer very high revisit rates of 1-5 days and high spatial resolutions from 3-10 m, respectively (Table 1), for multispectral imagery (Drusch et al., 2012;Butler, 2014;Breger, 2017). These high spatiotemporal resolutions open up unprecedented possibilities to study a wide range of landslide velocities and natural hazards through remote sensing. Continuing data access is fostered 85 by PlanetLabs and by Copernicus (via its open data policy) providing affordable or free data for research. Examples of landslide activity studies employing multi-temporal datasets based on this access to high spatiotemporal data include Lacroix et al. (2018), using Sentinel-2 scenes to detect motions of the 'Harmalière' landslide in France, and Mazzanti et al. (2020), who applied a large stack of PlanetScope images for the active Rattlesnake landslide, USA.
As landslides tend to accelerate beyond the deformation rate observable with radar systems before failure, we concentrate on 90 optical image analysis (Moretto et al., 2016). One advantage of optical imagery is its temporally dense data (Table 1) compared to open data radar systems with sensor repeat frequency every six days and revisit frequency between three days at the equator, about two days over Europe and less than one day at high latitudes (Sentinel-1, ESA). Optical data allows direct visual impressions from the multispectral representation of the acquisition target and the option to employ this data for further complementary and expert analyses. While active radar systems overcome constraints posed by clouds and do not 95 require daylight, data voids can be significant due to layover or shadowing effects in steep mountainous areas (Mazzanti et al., 2012;Plank et al., 2015;Moretto et al., 2016). Moreover, north/south facing slopes are less suitable, thus limit the range of investigation (Darvishi et al., 2018). In general, sensor choice depends on the landslide motion rate with radar at the lower and optical instruments at the upper motion range (Crosetto et al., 2016;Moretto et al., 2017;Lacroix et al., 2019).
However, a flexible, cost-effective alternative to spaceborne optical data are airborne optical images taken by UASs. Freely 100 selectable flight routes and acquisition dates enable avoiding shadows from clouds and topographic obstacles as well as unfavourable weather conditions and summer time snow cover, all of which frequently impair satellite images (Giordan et al., 2018;Lucieer et al., 2014). UAS-based surveys provide accurate very high resolution (few cm) orthoimages and digital elevation models (DEM) of relatively small areas, suitable for detailed, repeated analyses and geomorphological applications (Westoby et al., 2012;Turner et al., 2015). 105 In recent years, data provision for users has increased and today data hubs provide easy accessibility to rapid, pre-processed imagery. Nonetheless, technological advances can be misleading as they promise high spatiotemporal data availability, which frequently does not reflect reality (Sudmanns et al., 2019). One key problem is the realistic net temporal data resolution which is often significantly reduced due to technical issues, such as image errors and non-existent data (i.e. data availability, completeness, reliability). Other problems include data quality and accuracy in terms of geometric, radiometric 110 and spectral factors (Batini et al., 2017;Barsi et al., 2018). Knowledge of the most useful remote sensing data options is vital for complex, time-critical analyses such as ground motion monitoring and landslide early warning. Timely information extraction and interpretation are critical for landslide early warnings yet few studies have so far explicitly focused on time criticality and the influence of the net temporal resolution of remote sensing data.
In this investigation we propose both a conceptual approach to evaluating lead time as a time difference between the "time to 115 predict" and the "forecasting time" and assess the suitability of UAS sensors (0.16 m) and PlanetScope (3 m) imagery (the latter with temporal proximity to the UAS acquisition) for LEWS. For this we have chosen the 'Sattelkar', a steep, high-GeoEye-1 3 1.64 C Pléiades 1A/B 1 2.0 (0.5) 5 C PlanetScope 1 3.0/3.125 2 C/F 1 RapidEye 3 5.5 5 4 F Sentinel-2 A/B 5 10 F Landsat 8 16 30 F alpine cirque located in the Hohe Tauern Range, Austria (Anker et al., 2016). We estimate times for the three steps (i) collecting images, (ii) pre-processing and motion derivation by digital image correlation (DIC) and (iii) evaluating and visualizing. The results from the Sattelkar site -and from historic landslide events -will be discussed in terms of usability 120 and processing duration for critical data source selection which directly influences the forecasting window. Accordingly, we try to answer the following research questions: 1. How can we evaluate lead time as a time difference between the "time to predict" and the forecasting time for high spatiotemporal resolution sensors?
2. How can we quantify "time to warning" as a sequence of (i) time to collect, (ii) to process and (iii) to evaluate 125 relevant optical data?
3. How can we practically derive profound "time to warning" estimates as a sequence of (i), (ii) and (iii) from UAS and PlanetScope high spatiotemporal resolution sensors? 4. Are estimated "times to warning" significantly shorter than the forecasting time for recent well-documented examples and able to generate robust estimations of lead time available to enable reactive measures and evacuation? 130 2 Lead time -a conceptual approach 2.1. The conceptual approach Natural processes and their developments constantly take place independently, thus dictate the technical approaches and methodologies researchers can and must apply within a certain time period. For that reason, we hypothesise the forecasting window texternal is externally controlled, consequently the applicability of LEWS methods (tinternal) is restricted because they 135 must be shorter than texternal. This approach is the framework of our time concept (Fig. 1).

Figure 1
The novel conceptual approach for lead time, time to warning and the forecasting window for optical image analysis.
The forecasting window is started (texternal, green dashed outline) following significant acceleration exceeding a set displacement threshold, leading to a continuous process. Simultaneously with the forecasting window, time to warning 140 (twarning) starts (grey outline). Time to warning is divided into a three-phase-process to allow time estimations for a comparative assessment of different types of remote sensing data. This process consists of the phases (1) time to collect, (2) time to process and (3) time to evaluate, each with their individual durations. Confidence in the forecasted event increases with time as process acceleration becomes more certain. Once a warning is released (orange box), the lead time begins (tlead) and is terminated by the following release and subsequent impact (red box). The lead time is the difference between the 145 forecasting window and the time to warning. During the lead time, reaction time (treact) starts when appropriate counter measures are taken to prepare for and reduce risks ahead of the impending event, and ends with the final impact.
The time to warning period (twarning) is defined by the time necessary to systematically collect data, analyse the available information and to evaluate it. Hence, the greater the lead time, the more extensive countermeasures can be implemented prior to the event. An imperative for an effective EWS, the required time to take appropriate mitigation and response 150 measures has to be within the lead time interval (tlead) (Pecoraro et al., 2019) with tlead ≥ treact.

Practical implementation of multispectral data in the concept
The time to warning consists of a three-phase-process (see Sect. 2.1. and Fig. 1) to allow rough time estimations for a comparative assessment of different types of remote sensing data. Nevertheless, to realise this temporal concept an established, operating system is required, which includes reference data (DEM, previous results), experience from past field 155 work and ready UAS flight plans with preparation for a UAS flight campaign, satellite data access, experience in the single software processing steps including final classification and visualisation templates and, if utilised for UAS, installed and measured ground control points.
The first phase includes the collection of data starting from the acquisition by the sensor, the data transfer, image preprocessing and provision to the end user. The user selects images online from the data hub, downloads and organises them. 160 For a UAS campaign, the user must obtain flight permits, check flight paths and conduct the UAS flight. The second phase encompasses time to process for the complete data handling from the downloaded data to final analysis-ready image stacks in a GIS or a corresponding software. These preparatory steps may include image selection and renaming, atmospheric correction, co-registration, resampling and translation to other spatial resolutions and geographic projection systems, adjustments such as clipping, stacking of single bands into one multispectral image or the division into single bands, 165 calculation of hillshade from DEM among others, depending on the requirements. Following this preparation, the data is processed with the appropriate software tools to derive ground motion, calculate total displacement and derive surface changes, e.g. volume calculations or profiles. In the third and last phase, time to evaluate, the results are compared to inventory data and, if available, ground truth data, displacement results of other sensors or different spatial resolutions, different time interval variations to observe changes in sensitivity to meteorological conditions. Additionally, filters may be 170 applied to eliminate noise. Finally, the results are analysed and evaluated. In each phase quality management is carried out for data access and pre-and post-processing. In time to collect, the images must be selected manually prior to any download from the data hub, as its filter tool options on cloud and scene coverage are of limited help. Accordingly, the areal selection may be misleading as the region of interest (RoI) might not be fully covered, though the sought-for, smaller area of interest (AoI) is covered but not returned from the request. Concerning cloud filters, first, the filter refers to the RoI as a whole in 175 terms of percentage of cloud coverage. The AoI can still be free of clouds or else be the only area covered by clouds in the total RoI. Therefore, an image is either not returned although usable, or returned but not useable. Second, clouds can create shadows for which no filter is available. As a result, affected images have to be manually removed by the user. Images which are of low quality due to snow cover have to be discarded, too. These actions indirectly represent first quality checks in the collection phase. In the following processing phase, the images in a GIS, are checked for quality and accuracy. Depending 180 on the data provider, some pre-processing such as radiometric, atmospheric and/or geometric corrections may have been conducted. During this phase, additional user-based steps will be checked if necessary. Finally, the results are compared to other data (e.g. DEM, dGPS), reviewed for their validity and may be supplemented by statistical evaluation.

Study Site
The Sattelkar is a steep, high-alpine, deglaciated west-facing cirque at an altitude of between 2 130-2 730 m asl in the 185 Obersulzbach valley, Großvenedigergruppe, Austria (Fig. 2a). Surrounded by a headwall of granitic gneiss, the cirque infill is characterised by massive volumes of glacial and periglacial debris as well as rockfall deposits (Fig. 2b, c). Since 2003 surface changes have taken place as evidenced by a massive degradation of the vegetation cover and the exposure and increased mobilisation of loose material. A terrain analysis revealed that a deep-seated, retrogressive movement in the debris cover of the cirque had been initiated (Anker et al., 2016;GeoResearch, 2018). High water (over)saturation is assumed to be 190 causing the spreading and sliding of the glacial and periglacial debris cover on the underlying, glacially smoothed bedrock cirque floor forming a complex landslide (Hungr et al., 2014). Detailed aerial orthophoto analyses, witness reports and damage documentations indicate a steady increase in mass movement and debris flow activity over the last decade (Anker et al., 2016).

195
In August 2014, heavy ongoing precipitation triggered massive debris flow activity of 170 000 m³ in volume, of which approximately 70 000 m³ derived from the catchment above 2 000 m. A further 100 000 m³ was mobilised in the channel 200 within the cone. The consequence was that the Obersulzbach river was blocked leading to a general flooding situation in the catchment, resulting in substantial destruction in the middle and lower reaches (Fig. 3).  In the Sattelkar cirque, several monitoring components are installed to provide ongoing and long-term monitoring. Nine permanent ground control points (GCPs) measured with a dGPS to provide stable and optimal conditions to derive orthophotos from highly accurate UAS images (GeoResearch, 2018). A total number of 15 near surface temperature loggers (buried at 0.1 m depth) recorded annual mean temperatures slightly above the freezing point (1-2 °C) in the period 2016 to 215 2019. Ground thermal conditions at depth react with significant lag times to recent warming and therefore are primarily determined by climatic conditions of the past (Noetzli et al., 2019). Significantly cooler climatic conditions in previous decades and centuries (Auer et al., 2007) thus likely contributed to the formation of (patchy) permafrost at the Sattelkar cirque. Recent empirical-statistical modelling of permafrost distribution in the Hohe Tauern Range confirms possible permafrost presence at the study site (Schrott et al., 2012). 220 The Sattelkar is a suitable case study as it is in the early stages of the landslide development and thus fits best to this conceptual approach. Here, processes take place on time scales appropriate for long-term observation to provide sufficient warning time. The active part of the cirque has accelerated in recent years allowing the analysis of EWS concepts based on multispectral optical remote sensing data supported by complementary block tracking. 225

Optical imagery
Optical satellite imagery is more appropriate for high deformation studies than radar applications due to the high spatial resolution as well as the short time span between acquisitions (Delacourt et al., 2007). Although the west-facing slope is favourable for the application of radar derivatives (InSAR/DInSAR), the choice to use optical imagery is based on the 230 observed high displacement rates, which cause decorrelation when using radar technologies as they are more sensitive than optical technologies. Complex and/or large displacement gradients make the phase ambiguity difficult to solve for radar interferometry (Kääb et al., 2017). Revisit times of current radar satellites (e.g. Sentinel-1) are longer than those of optical satellites, and if time periods between image acquisition become too long, ground motion may accumulate such that the displacement is too high to be measured. Several studies on displacements of faults and landslides have shown the potential 235 of optical data to provide detailed displacement measurements based on image correlation techniques (DIC) (Leprince et al., 2007;Rosu et al., 2015). A further advantage of optical images for geomorphological processes in steep terrain is their viewing geometry (close to nadir) . Here we employ DIC to compare the spatiotemporal resolution of multispectral optical imagery (UAS and PlanetScope) and to assess its suitability for early warning purposes. UAS images offer excellent spatial resolution and accuracy at the centimetre scale (Turner et al., 2015) and complement large scale 240 satellite or airborne acquisitions (Lucieer et al., 2014). PlanetScope imagery provides the highest temporal resolution among available sensors with daily acquisitions, guaranteed data availability, and free and open access for research purposes. In this study the PlanetScope Analytic Ortho Scene SR (surface reflectance) imagery (16-bit, geometric-, sensor-and radiometric corrections) was employed (Planet Labs, 2020b) and was supported by the Planet Labs Education and Research Program.

Data availability of PlanetScope 245
Research on the availability and usability of PlanetScope imagery was conducted on the Planet Explorer data hub for the time span from the beginning of April to the end of October in 2019, as during these months snow cover should be negligible. Filter parameters were solely set for 4-band PlanetScope Ortho Scenes and the Sattelkar AoI. In order to obtain all available images, no filters (e.g. sun azimuth, off nadir angle) were applied. We defined four categories i) meteorological constraints due to snow cover, cloud cover and cloud shadow; ii) image (coverage) errors made by the provider, iii) no data 250 availability and iv) the remainder of usable data ( Table 2). The output request was evaluated according to the defined categories and was compared to the provider's guaranteed daily image provision, which is comprised of 213 days for the time period (01.04.2019-31.10.2019). We calculated percentages for the above categories based on days per month as well as a seven-month sum and percentage average. The availability analysis did not include an examination of the data with regard to its spatial usability: positional accuracy and/or image shifts. 255 respectively, on all 213 days; on average 14.5 % and 7 % of the days were not usable (Table 2). For 10 days in June snow influence had the greatest negative share (33.3 %), for April there were three days of snow coverage and the months September and October each had one day of snow coverage. Cloud cover/shadow exerted a higher impact on data usability by 14.5 %. Problems on the part of PlanetLabs made much of the data unusable due to image errors; between four and nine images per month were not usable (21 %). On average for 26.2 % of the analysed time period no image data was available. 265 In this seven-month period, 43 images (20.1 %) had data voids or did not cover the AoI, thus the overall usability is limited to about 11 %.

Data Acquisition and Processing
In line with the concept in Fig. 1 (Sect. 1), the following processing steps are categorised and described.
(1) tcollect: UAS data acquisition was preceded by detailed flight route planning and checks of local weather and snow 270 conditions. UAS flights were carried out with a DJI Phantom4 UAS on 13.07.2018, 24.07.2019 and 04.09.2019 (see Table 3, Fig. 4, Fig. 6b, c). For each acquisition, the total area was covered by four flights which were started on different elevations (Table 4). Flight 275 planning was done with UgCS maintaining a high overlap (front: 80 %, side: 70 %) and a target ground sampling distance (GSD) of 7 cm. The area covered was approximately 3.4 km² and with a flight speed of about 8 m/s total flight time took 3.5 hours. The images were captured in RAW format. In the Planet Explorer Data Hub, PlanetScope Ortho Scenes were selected for usability; imagery affected by snow cover, cloud cover, cloud shadow and partial AoI coverage was discarded (Table 5). 280    The UAS images in RAW format were modified using Adobe Exposer to improve contrast, highlights, shadows and clarity.
Thereafter, they were exported as JPG (compression 95 %) and processed with Pix4Dmapper to 0.08 m resolution and 300 orthorectified based on nine permanent ground control points (GCP, 30 x 30 cm). These were repeatedly (1000 measurements/position) registered with the TRIMBLE R5 dGPS and corrected via the baseline data of the Austrian Positioning Service (APOS) provided by the BEV (Bundesamt für Eich-und Vermessungswesen). Horizontal root-meansquared errors (RMSE) range from 0.05 m to 0.10 m for vertical RMSE. These GCPs were employed for georeferencing and further rectification of all UAS surveys. 305 Next, the data was clipped to a common area of interest (AoI) and resampled with GDAL and the cubic convolution method to 0.16 m to enhance processing time and increased reliability of image correlation. PlanetScope Satellite images were coregistered in Matlab relative to a reference image (https://gitlab.lrz.de/tobi.koch/satelliteregistration.git). A feature point detection step was applied to estimate a geometric similarity transformation between the reference (master) and all target (slave) image pairs excluding the AoI with its terrain motion. Thereafter feature point outliers were statistically removed 310 (RANSAC) and the similarity transformation of the slave images to the master image was performed. After removing the outliers, more than 500 feature matches were found for the entire image pair dataset. The mean distance of transformed inlier feature points from the target image to their corresponding feature matches in the reference image ranged between 0.6 and 0.8 pixels, confirming the high registration accuracy (see OSM Fig. 14). We used digital image correlation (DIC) to measure the displacement for the active landslide body of the Sattelkar and to assess the suitability of the PlanetScope and UAS data. 315 This method employs optical and elevation data and calculates the distance between an image pair, based on the spatial distance of the highest correlation peaks between an initial search and a final reference window. The result provides displacement and ground deformation in 2 D on a sub-pixel level. COSI-Corr (Co-registration of Optically Sensed Images and Correlation), a widely used software in landslide and earthquake studies was used for sub-pixel image correlation (Stumpf, 2013;Lacroix et al., 2015;Rosu et al., 2015;Bozzano et al., 2018). COSI-Corr is an open source software add-on 320 developed by CALTECH (Leprince et al., 2007), for ENVI classic. There are two correlators; in the frequency domain based on FFT algorithm (Fast Fourier Transformation) and a statistical one. Applying the more accurate frequential correlator engine, recommended for optical images, different parameter combinations of window sizes, direction step sizes and robustness iterations were tested. Parameter settings include the initial window size for the estimation of the pixelwise displacement between the images and the final window size for subpixel displacement computation in x, y; a direction step 325 in x, y between the sliding windows; and several robustness iterations (Table 6). We utilised recommended window sizes as suggested by Leprince et al. (2007) and Bickel et al. (2018).
Step size one showed good results while keeping the original spatial resolution for the output; robustness iterations of two to four were sufficient for our purposes. Initial and final window sizes were systematically tested (see Table 6). For computing a state-of-the-art powerstation was employed (AMD Ryzen 9 3950X 16-core processor, 3.70 GHz, 128 GB RAM). 330 The results of each correlation computation returns a signal-to-noise ratio map (SNR) and displacement fields in east-west and north-south directions. These results were exported from ENVI classic as GTiff and the total displacement was then 335 calculated with QGIS.
(3) tevaluate: In the last phase (time to evaluate) the results of various parameter settings were compared in QGIS and ArcGIS along with different combinations of visualisation. Displacement below a 4 m threshold was discarded from the PlanetScope datasets due to aberrant values (noise, outliers). The threshold definition was defined on (i) the value distribution in both the total displacement and the corresponding SNR result, and (ii) a visual comparison of the maps for the total displacement and 340 the SNR. This definition allowed us to identify outliers and unlikely displacement. Apart from this threshold no other filters were employed, and we kept the output raw (see for raw DIC on PlanetScope OSM Fig. 13). Very few inconsistencies were present in the UAS-derived displacement results, which were accepted without modification.
Additional analyses were performed to estimate the DIC outputs of both, the UAS orthophotos and PlanetScope satellite imagery. Visual tracking of 36 single blocks, identifiable in the UAS orthophoto series allowed deriving direction and 345 amount of movement; this supported the confirmation process for (i) the total displacement and (ii) the results of automated and manual tracking. In the next section we present this approach only for time interval II.

Results
In Sect. 5.1. we present ground motion results from DIC for the original input resolution for i) UAS, 0.16 m and ii) PlanetScope, 3 m input resolution based on parameters in Table 6. In Sect. 5.2. DIC results for UAS, 0.16 m are analysed 350 with regard to displacement of visual single block tracking. Finally, in Sect. 5.3. required times for tcollection, tprocessing and tevaluation for each sensor are presented.   Fig. 5b show the total displacements derived from UAS orthophotos at 0.16 m resolution for time intervals I 360 and II (see Table 6). Apart from several minor displacement patches, no motion is visible outside the active body in either period. Time interval I (376 d) (Fig. 5a) shows mean displacement values from 6 to 14 m for a coherent area in the eastern half of the lobe from the centre (c) to the eastern boundary of the active area. The highest displacement rates (up to 20 m) are observed within small high-velocity clusters in the northwest section (d). Lower velocities occur along the southern boundary (e, f), ranging from zero to 6 m with smooth transitions. Ambiguous, small-scale patterns with highly variable 365 displacement rates are present in the western half (a) and along the northern boundary (b). No motion is detected along the western fringe (i.e. at the landslide head) which is 20 m in width. South of the landslide (g) there is a small patch of minor displacement with continuous (up to 3.5 m) and ambiguous signals. Furthermore, we observed small-scale patterns of ambiguous signals in the east (j) and in the west of the active area in the drainage channels (h, i).

Total displacements
Time interval II (42 d) (Fig. 5b) shows great similarity to time interval I with ambiguous signals in the same areas such as 370 the drainage channels (h, i) and within the western half of the active area (b). In contrast to interval I (Fig. 5a), within the active area a homogenous higher velocity patch (up to 6 m) near the landslide head is evident (a). In the eastern half large homogenous patches extend from the landslide centre (c) to the root zone (d) showing coherent displacement values of zero to 4 m. During this shorter time interval II, no displacement is detected along the south eastern boundary (e) and for large parts of the root zone (f) previously covered in I. Similar to I, the landslide head has a 20 m rim free of signal (also see Fig. 6  375 x, y). In the central part of the lobe (c) total displacements are significantly reduced.  Table 3 and Fig. 4). For interval Ib (370 d) (Fig. 5c) wide fringes with no motion were detected around an actively moving core area, which consists of small-scale clusters with variable total displacement in the western part, coherent high velocities in the middle, and coherent low velocities east of this core area. Outside the landslide, northeast and immediately south (j), high-velocity 380 patches are observed.
In interval II (42 d) (Fig. 5d) the detected displacement is restricted to the western half of the landslide (a) and shows the same significant fringes with no motion as in I. Compared to interval I the motion pattern of this core area is more homogeneous with increasing displacement towards the east. Outside the active area several patches show medium to high total displacement, the largest of which is located 300 m northwest of the landslide (i). 385

Time required for collection, processing and evaluation
In Sect. 2 we introduced a novel concept to extend lead time, consisting of three phases within the warning time window (see 405 Fig. 1). This concept is based on DIC results, thus every step comprised in each phase has been previously undertaken. On this basis, knowledge of required time for a further process iteration of the three phases is given.
Time required for collection, processing and evaluation of UAS and PlanetScope data are estimated and summed in Fig. 7.
PlanetLabs specifies 12 hours from image acquisition to the provision in the data hub, which includes to a large amount data pre-processing (Planet Labs, 2020b). Adding two hours for the selection, order and download process, we assume that time 410 required for the collection phase is approximately the same for both sensors, with 14 hours for PlanetScope and 12 hours for UAS. With regard to the time needed for the processing phase, the sensors differ with UAS requiring 17 hours and PlanetScope five hours. Time for the evaluation phase is estimated to be about two hours. In sum, twarning for UAS is approximately 31 hours compared to 21 hours for PlanetScope.

Figure 7
Time to warning is composed of three phases: time to collect, to process and to deliver. Time to warning (subsequent to acceleration) is 21 h for PlanetScope and 31 h for UAS. Thus, any hazard process that takes longer than 21/31 h to prepare the release and impact can be forecasted.

Discussion
To systematically analyse the predictive power of the UAS and PlanetScope data, we will (i) evaluate ambiguous signals, 420 error sources and output performance, (ii) assess obtainable temporal and spatial resolution and (iii) derive a systemic estimate of the minimum obtainable warning times.

Error sources and output performance
To evaluate error sources and output performance, we compared results of digital image correlation results from optical data with (i) high resolution UAS orthophotos, (ii) mapped mass movement boundary and (iii) visual block tracking for UAS 425 orthophotos. The approximately one year evaluation period encompassed all seasons, hence freezing/thawing conditions and a wide range of meteorological influences, e.g. thunderstorms and heavy rainfall, are included. The two investigated time intervals are I/Ib and II, covering 376/370 days and 42 days (typical high-alpine summer season), respectively (Fig. 4).
Interval II exclusively covers (high-alpine) summer conditions, with negligible to no contribution from freezing conditions. As these inclusion periods are inconsistent, the amount of total displacement cannot be directly compared; however the 430 relative motion patterns can be. Accordingly, we can confirm the suggested parameter settings of earlier studies on window sizes, steps and robustness iterations (Ayoub et al., 2009;Bickel et al., 2018).
In terms of the mass movement boundary, the total displacement derived from the DIC of the UAS data generally matches the field-mapped landslide boundary for both intervals (I, II) (Fig. 5a, b), and is supported by the absence of significant noise outside the AoI. Mapped boulder trajectories for interval II (see Fig. 6) are consistent with the calculated total 435 displacement and thus confirm COSI-Corr as a reliable DIC tool to derive ground motion for this study site and UAS orthophotos as suitable input data. Nevertheless, there are several areas with ambiguous signals. Here we follow Leprince et al. (2007) describing a correlation loss as 'decorrelation' with signal-to-noise values of low/null (i.e. no convergence of the correlation algorithm) and/or large offsets, either unrealistic in nature or beyond the valid matching window distance. Decorrelation in our understanding exhibits a salt-and-pepper appearance in the DIC result with random displacement 440 vectors, related to inconsistently tracked features. The software is not able to find the corresponding, correlated surface pattern, leading to a misfit (i.e. misrepresentation) and/or mismatch (i.e. blunders) of the matching windows and finally resulting in noise (Debella-Gilo, 2011;Guerriero et al., 2020). Nevertheless, this decorrelation signal is still a valuable observation that might be related to surface processes and not only to erroneous limitations of the DIC method. There are three main reasons that might cause these effects: (i) significant temporal change of the surface, i.e. revolving and/or 445 rotational deformation, (ii) high displacements exceeding the matching window size being smaller than the offset, (iii) land cover changes such snow cover, vegetation cover and alluvial processes, among others, and (iv) changes related to illumination (e.g. shadow) or image errors (e.g. orthorectification, shifts in individual bands) (Leprince, 2008;Debella-Gilo, 2011;Lucieer et al., 2014;Stumpf et al., 2016). In our study, the decorrelated salt-and-pepper areas include to a large degree the landslide head (a), the drainage channel (h) (Fig. 5a, b), a larger patch south of the active area boundary (g) 450 (Fig. 5a), and some smaller ones in little depressions (g) (Fig. 5a) and (j) (Fig. 5a, b). The patches (j) and east of (j) are identified as snow fields in the orthophotos and the noise results from decorrelation. In Fig. 5a, the large southern patch (g) shows clear displacement values for the rear part and decorrelation for the front region resulting from morphological changes within the image pair of interval I (see OSM Fig. 12). This is due to a gain between 1 and 2 m for an area of about 250 m².
The decorrelation in the drainage channel (h) could stem from massive changes in pixel values, similar to the decorrelation 455 on the basis of alluvial processes, as described by Leprince et al. (2007). Decorrelations in the areas with the fastest ground motions also lead to high pixel changes (Stumpf et al., 2016): these are observable in the active landslide area within the lobe, where large areas of decorrelation may be explained by high displacements in the leading landslide head (a) with redetected, hence correlated pixels in the trailing areas (c, d, e, f). These findings can be transferred to the landslide interior area (a, b), the frontal western regions and the northern margin (b). The observation is confirmed by geomorphological 460 mapping (see OSM Fig. 11) and measured boulder block trajectories from the orthophotos (Fig. 6a). Several patches of correlation (c, f) with corresponding boulder trajectories up to 4 m (34.8 m yr -1 ) (d) can be detected in the rear areas. A correlated patch with a 16 m (34.8 m yr -1 ) trajectory (a) is located in the flow direction behind the foremost boulder. In this case the method was able to partially capture the displacement as the distinct boulder block supported the detection, which probably led to correlation. Similarly there is another example with a trajectory of 10 m (86.9 m yr -1 ) outside a homogeneous 465 correlated area. This leads to the assumption that for the calculated time period, with 63 pixels or more at a resolution of 0.16 m, no pixel matching is possible and probably reached the correlation capacity due to the too high displacement. With a correlation window smaller than the displacement, the algorithm cannot capture the displacement (Stumpf et al., 2016).
However as field observations provide evidence that the rock masses are deforming, and the surface is altering due to the high mobility and rotational behaviour of some boulder blocks. This leads to changed pixel values and spectral 470 characteristics of the block surface and the surrounding area, which can also result in poor correlations, and even random errors and mismatches (Debella-Gilo and Kääb, 2011). This finding is similar to observations in a rock glacier study by Debella-Gilo and Kääb (2011). Similar results were observed by Lucieer et al. (2014), who described a loss of recognisable surface patterns if revolving and rotational displacements occur, causing decorrelation and a noise as output. These results show that with COSI-Corr and UAS orthophotos of 0.16 m, it is possible to detect the total displacement of the landslide in 475 both extent and internal process behaviour even in this steep, heterogeneous terrain. Nevertheless, high displacement rates and rotational surface behaviour in the cirque limit the DIC method. A decrease of the time interval for this particular highly mobile study site would likely reveal an enhanced correlation since for shorter time periods the total displacement decreases, and surface changes are reduced, which can be controlled by shortening the temporal baseline (Debella-Gilo and Kääb, 2011). In sum, though the results contain heterogeneous, noisy, decorrelated areas, the combination with homogeneous 480 displacement areas still offers valuable insights into this and other internal landslide structures and complex behaviours.

Comparison of temporal and spatial resolution
We compared the COSI-Corr total displacement results of PlanetScope (Ib and II, Fig. 5c, d) and UAS images (I and II, Fig. 5a, b and Fig. 6a) for the same time periods at different spatial resolutions (see Table 6). For the PlanetScope DIC result the main part of the landslide is detected, and its area is generally consistent with the results of the UAS DIC, which is 485 additionally confirmed by boulder trajectories. The frontal part (a) reveals correlation signals (I and II); while for the same time intervals and parts, the UAS DIC results show a decorrelation (Ib and II). The correlation is likely to be attributable to the coarser spatial resolution of 3 m PlanetScope input data, hence a smaller number of pixels to be captured at this site with the DIC method. Similar texture of rock clast surfaces could lead to false positives resulting in correlation as patches appear similar in matching windows. However, in contrast to the UAS result (Fig. 5a, b), the outcome on a large scale fails to detect 490 the entire actual active area (b), (f) as well as its internal motion behaviour. Nevertheless, for the visualisation and analysis of the PlanetScope results, the range of total displacements had to be restricted to values equal to and greater than 4 m due to noise and outliers over large areas, as applied and described by Bontemps et al. (2018). Even then, noise and several misrepresented displacement patches are observed for (i, j) and in the northeast image corner (Fig. 5). We can identify several reasons for these large clusters of high motion values. Massive cloud and snow coverage hampered both first images 495 of interval Ib (19.07.2018) (Fig. 5c) and II (24.07.2019) (Fig. 5d), leading to a 20 m fringe of false displacements in the north-eastern part of the image. Minor snow fields as visible in the images from 24.07.2019 for both, the UAS and PlanetScope, likely explain the big cluster of incorrect displacement southeast of the lobe (j); nonetheless, in the satellite image they are smaller than the resulting DIC displacement. High cloud coverage in two input images with large areas of white pixels may exert an influence leading to high gains due to sensor saturation (Leprince, 2008). Illumination changes in 500 interval II (Fig. 5d) may cause unrealistic displacements outside the boundary with slightly darker colours due to shadows in the first satellite image (24.07.2019) and large parts within the second image (04.09.2019) are also in the shade. A comparison of the acquisition times and true sun zenith, e.g. for the second image, reveals a difference of 01:34 h between the image acquisition at 11:36 LT (local time) and the true local solar time at 13:10 LT. As the study site is located in a high-alpine terrain with a west facing cirque, at this time of day there are shadows of considerable length which have a 505 significant influence on the result of digital image correlations. One clear advantage of the UAS images is that their acquisition is plannable according to the best illumination conditions with the sun at its zenith. Moreover, the UAS flight path as well as the system itself remained the same for all three acquisitions, while PlanetScope employs various satellites.
Despite different input resolutions and time intervals (Ib vs. I and II vs. II, see Table 3) with different sensors (UAS, PlanetScope) there is a similarity for the landslide head which indicates that the displacement is restricted to a smaller area 510 than the previously demarcated boundary, based on our field investigations. This is clearer for the time interval I (376/370 d) (Fig. 5a vs. c) as for the longer temporal baseline the total displacement accumulation is higher, thus better captured by COSI-Corr for PlanetScope 3 m resolution. Due to the shorter interval II (42 d) (Fig. 5b vs. d) with less accumulated total displacement, the rear of the landslide is not represented; no signal is shown as the total displacement for PlanetScope was restricted to values above 4 m. Values below 4 m had to be discarded for PlanetScope DIC results as they were lost in noise, 515 i.e. for the entire DIC results there is total displacement between 0 m and 4 m (cf. the online supplementary material (OSM) Fig. 13). Hence, when applying a minimum threshold of 4 m, the satellite image detects large parts of the main active core area but widths of 50-80 m from the boundary show no displacement. However, false large clusters of high total displacement are within the PlanetScope result interval I for (j) and the northeast image corner (Fig. 5c), and interval II for (i) (Fig. 5d). 520 Measured ground motion of block tracking and PlanetScope results indicate and support existing high ground motions. In addition there are morphologically significant volumetric turnovers with areas of large gains and losses between ± 5 m (see OSM Fig. 11). These observations might explain the resulting decorrelation at the finer resolution of 0.16 m for the landslide head: the matching window is smaller than the offset and texture surface changes are too complex to be re-detected, i.e. matched, and thus correlated, leading to decorrelation and noise. Homogeneous correlated patches are in the front of the 525 landslide body for the shorter time interval; there may have been some displacement just below the detection threshold for this high ground motion or some boulders and their surroundings might have been matched, or both ( Fig. 6a (a)). In this case for the complex ground motion with high spatial resolution data, the previous assumption based on a shorter time interval likely leads to improved detection of inherent process behaviour (see Sect. 6.1.). Generally, with high resolution images, such as UAS, we recommend first calculating displacements based on a coarser input resolution (1-3 m) to examine the 530 overall situation and detect changes, and second to calculate displacements at a finer resolution in order to focus on relevant details of the AoI. With regard to PlanetScope data, a 3 m resolution seems to be in a good spatial range to assess ground displacements even of this steep and heterogeneous study site with its high motion. Nonetheless, constraints such as illumination due to early daytime acquisitions leading to shadows, meteorological influences by clouds, cloud shadows and snow decrease the quality of the satellite images and reduce their applicability. Sensor saturation, shadow length, size and 535 direction as well as changes in snow, cloud or vegetation cover impose limitations (Delacourt et al., 2007;Leprince et al., 2008) and accord with our observations. The authors identify additional limitations such as radiometric noise, sensor aliasing, man-made changes and co-registration errors (ibid.). All these limitations have a negative impact on the input image, which leads to impaired DIC calculations and results, and (partially or wholly) inaccurate analysis of the displacement. These might have played a role in our results. In our experience, the usability of the DIC result may be 540 influenced by the input image quality. This restricts the application of PlanetScope images to a certain degree. They can be employed as input data to detect displacements, but as there are in the present setting too many signals of false-positive displacements, which can solely be discarded on the basis of field evidence, this data is currently of limited use. It should be handled with caution and we recommend combining it with complementary data and ground truth.

Estimating time to warning 545
Early warning is essentially defined as being earlier than the event and thus puts high external time constraints on observation and decision. The time window between the detection of an accelerating movement preparing for final failure and the final failure itself is determined by the environment. Therefore, two sensors with the highest available spatiotemporal resolution were evaluated and compared with regard to their applicability to the early warning of landslides. We made rough assumptions and assessed the time needed for the phases of time (i) to collect, (ii) to process, and (iii) to evaluate relevant 550 data (summarised in the time to warning window, see Fig. 7).
Despite different underlying technologies the time required for the collection phase is approximately the same for both sensors. For UAS, we estimated about 12 hours under ideal circumstances, while for PlanetScope 12 hours (Planet Labs, 2020b) plus two hours for image selection, download and initial analysis, adding up to 14 hours in total (see Sect. 5.3.). In the second phase, time to process, deriving orthophotos from raw UAS images is time consuming. The subsequent DIC 555 calculations demand significantly more processing time for the UAS images than for lower resolution PlanetScope images.
The final phase, time to deliver, takes about two hours for each sensor. In our case study, the estimated time to warning (twarning) was 10 h longer for the UAS approach (31 h) in comparison to the Planet Scope approach (21 h). These time calculations are based on ideal environmental conditions and data availability. Assuming good conditions exist to conduct the UAS flight and no constraints limit the utilisation of satellite images, in theory a daily deployment is possible. In reality, 560 unfavourable weather conditions, cloud and snow cover as well as limited data availability will increase the actual twarning significantly. From the available images in the Planet Data hub (besides other exclusions) meteorological influences reduced for April-October 2019 the usability by 14.5 % and 7 % for cloud cover and snow cover, respectively ( Table 2). The flexibility of a UAS can serve as a practical remote sensing tool for the investigation of ground motion behaviour in a spatiotemporal context. Nonetheless, weather influences can make a UAS flight impossible or impractical as the result might 565 be useless. Depending on the level of illumination, the same may apply for satellite images. Regardless of any meteorological constraints, the promised daily availability by PlanetScope is unrealistic, due to data gaps and provider issues; our study showed that for the Sattelkar from April to October 2019 only 11 % of the captured images during this time were usable. Hence, PlanetScope data has a temporal availability similar to Sentinel-1 with a 6-day revisit time. In timecritical early warning scenarios, when time is running out, all available even partly usable images will be utilised and 570 fieldwork may be conducted, even if the prevailing conditions are suboptimal but will increase data availability. The comparison of two selected remote sensing options demonstrates that the comprehensive knowledge on the available remote sensing data sources and their respective time requirements can substantially reduce the time to warning (twarning) and to extend the lead time (tlead).
Significant observations of the temporal evolution of historic landslides are presented in Table 7 and described below. These 575 include (i) the Preonzo rock slope failure, CH (Sättele et al., 2016;Loew et al., 2017), (ii) the Vajont rock slide, ITA (Petley and Petley, 2006) and (iii) the Sattelkar complex slide, AUT (Anker et al., 2016). These landslides have specific evolution histories, e.g. early observed crack developments, increased movement and minor events like Preonzo (2002 and 2010) (Sättele et al., 2016); Sattelkar, with large volume mass wasting processes since 2005 and a debris slide event in 2014 (see Sect. 3 Study Site) (Anker et al., 2016); and Vajont, with ductile failures in 1960 and 1962 and a transition from ductile to 580 brittle behaviour in 1963 (Petley and Petley, 2006;Barla and Paronuzzi, 2013). Table 7 Relevant dates for historic failures of Vajont (ITA), Preonzo (CH) and Sattelkar (AUT). Time period in italics-bold used for Fig. 9. Time intervals in days (~ for rough estimations) and years in square brackets; sum of days based on the first day of the month, if only month as reference is available from literature (Petley and Petley, 2006;Anker et al., 2016;Sättele et al., 2016;Loew et al., 2017).

585
Further explanation below. Figure 9 is the extension of our concept (see Sect. 1, Fig. 1) systematically supplemented with our estimated time to warning (UAS, PlanetScope), and compared to the few data series predating larger slope failures.
Following a significant acceleration, the forecasting window is opened and twarning starts, which is composed of phases (i) 590 time to collect, (ii) time to process and (iii) time to evaluate. To ascertain a significant acceleration one further observation is required. Hence, one complete cycle of the three phases, previous analyses and processing iterations are given. Our analysis showed that UAS and Planet Scope can approach times as short as 31/21 h, as a result tlead is increased and so is treact. Figure 8 Conceptual approach with estimated twarning for UAS and PlanetScope. Phases of collection, processing and evaluation (indicated as arrows of relative length in orange, blue and green, respectively) (see phases in Fig. 1 and Fig. 7) with their total duration time (grey dashed arrows). In twarning, one additional observation requires in sum 31 h for UAS and 21 h for PlanetScope data. Above, major landslides are compared from the onset or displacement detection (solid line) (Petley and Petley, 2006;Anker et al., 2016;Sättele et al., 2016).

600
Assuming both sensors reliably estimate ground motion, solely based on their time requirement, this concept was applied to the temporal development of historic landslide events, thus from measured increased displacements and/or massive accelerations to the final event (Table 7). On this basis we simplified the graph and what we defined as "significant acceleration" using dates of observations such as increased crack opening (Vajont), critical displacement (Preonzo) and the beginning of active ground motion (Sattelkar). Therefore, the opening of twarning and forecasting window are concrete 605 observations of the particular site, independent of any intensity described by the corresponding authors and allows more freedom for temporal evaluations without going into details.
For the Preonzo case, the entire 2012 spring period was characterised by high displacement rates. We defined the first of May 2012, when geologists operating the warning system informed local authorities and assembled a crisis team, as the onset or 'increased movement' and the 15.05.2012 with 300 000 m³ as the impact (Sättele et al., 2016), in total 610 approximately 15 days. For Vajont, the 1/velocity plot by Petley and Petley (2006) (based on data from Semenza and Ghirotti (2000)) shows an increase in movement at about day 60 along with a transition from a linear to an asymptotic trend at approximately day 30, defined as a transition from ductile to brittle. Therefore, we assumed 30 days of forecasting window for twarning and tlead until the impact of the hazardous event on 09.10.1963. However, note that velocities of about 35 mm d -1 are still low and at the minimum of the displacement recognition capability for the digital image correlation method. 615 For the Sattelkar site, the observed mass displacement increase is presumed to have started in 2005 with the 170 000 m³ debris flow event on 31.07.2014 as the impact, thus about 3 498 days (Anker et al., 2016).
Even for the Preonzo event, with its short forecasting window of 15 days, the ground motion assessment based on the evaluated optical remote sensing images, would have been possible under the assumption of reasonably good UAS flying conditions and the provision of usable PlanetScope images. For twarning there is enough temporal leeway to repeat at least 620 three to four successive measurements comprising the three phases. However, as single accelerations are possible in very short time intervals of less than two days, it is impossible to capture these accelerations by means of optical remote sensing methods, given a time requirement of 31 hours for UAS and 21 hours for PlanetScope. Nevertheless, this comparison shows that for larger and long-preparing slope failures the technical twarning may well be shorter than the forecasting window starting at the time at which the process becomes predictable. For this type of slope failures recent developments such as the ESA's 625 Geohazards Exploitation Platform (GEP), developed and operated by Terradue, supports on-demand services such as the Thematic Exploitation Platforms (TEPs) and has the potential to decrease twarning: The service provides an archive of Copernicus' Sentinel-1 and -2, Pléiades and Spot 6/7 data, and access to cloud computing resources to support large scale geohazard mapping and monitoring (Volat et al., 2017;Foumelis et al., 2019;Lacroix et al., n.d.). Therefore, the time critical phases of time to collect and time to process, which in our example are attributed to the larger share of the total time 630 requirement for twarning, could be significantly reduced as the data is directly accessible through high performance cloud computing. What remains is the third phase, time to evaluate, where a relatively short time is required, thus tlead is extended.

Conclusions and outlook
This paper presents an innovative concept to compare the lead time for landslide early warnings, of two optical remote sensing systems. We tested this temporal concept by applying UAS and PlanetScope images of temporal proximity as these 635 are currently the sensors with the best spatiotemporal resolution. We assessed the sensors' capability to identify hot spots and to recognise behaviour by delineating ground motion employing digital image correlation (DIC). In so doing, knowing the necessary processing time enabled us to estimate the time requirement and finally to incorporate it into the concept to evaluate sensors with regard to ongoing landslide processes of the Sattelkar as well as historic landslide events.
Our findings derived from DIC for this steep high-alpine case study show that high resolution UAS data (0.16 m) can be 640 employed to identify and demarcate the main landslide process and reveal its heterogeneous motion behaviour as confirmed by single block tracking. Thus, validated total displacement ranges from 1-4 m and up to 14 m for 42 days. PlanetScope Ortho Scenes (3 m) can detect the displacement of the landslide central core, however, cannot accurately represent its extent and internal behaviour. The signal-to-noise ratio, including multiple false-positive displacements, complicates the detection of hotspots at least in this very steep and heterogeneous alpine terrain. 645 Coarse temporal data resolution, such as in the case study investigated here, represents an important restriction to the use of optical remote sensing data for landslide early warning applications. Acceleration (and the resulting failure) over short periods of time will likely go unnoticed due to large data acquisition intervals. However, for prolonged acceleration periods, such as observed at the Sattelkar slide and many other relevant hazard sites, the chosen data sources have been demonstrated to represent a formidable early warning approach capable of contributing to an improved risk analysis and evaluation in 650 steep high-alpine regions.
With regard to the temporal aspect for early warning purposes, PlanetScope satellite images require less time compared to UAS for the time phases of collection, processing and analysing. As a consequence, when time is of the essence, the UAS acquisition cannot compete with the high frequency of PlanetScope daily revisit rates. In general, both are limited in their use as they are passive optical sensors dependent on favourable weather conditions. Nevertheless, with a realistic 10 % of 655 usable data for our study site, PlanetScope cannot provide daily data as promised.
To conclude, in methodological terms DIC is a reliable tool to derive total displacement of gravitational mass movements even for steep terrain. Given the high reliability of UAS data, its temporal resolution is the key in future attempts to overcome decorrelation due to high ground motions. In addition, a slightly coarser resolution reduces the time needed for total processing, enhances correlation while maintaining spatial accuracy and reliability. PlanetScope is especially interesting 660 as a complementary sensor when UAS employment is restricted e.g. inaccessible and/or dangerous sites or for areas too extensive to be covered. For continuous monitoring and early warning, the warning time window could be shortened by onsite drone ports with autonomous acquisition flights and automatic processing. Our systematic evaluation of the sensor capability can be applied to other optical remote sensing sensors, and the same is true for our conceptual approach which extends the lead time. Future studies should focus on the applicability of complementary optical data to confirm the 665 detection of landslide displacement and adjust UAS output resolution as this significantly increases the validity of DIC internal ground motion behaviour.

Data availability
PlanetScope data are not openly available as and PlanetLabs Inc. is a commercial company. However, scientific access 670 schemes to these data exist.

Author contribution
Doris Hermle developed the study together with Markus Keuschnig and Michael Krautblatter, analysed the data and wrote