Development of smart boulders to monitor mass movements via the Internet of Things: A pilot study in Nepal

1 Boulder movement can be observed not only in rock fall activity, but also in association with other 2 landslide types such as rock slides, soil slides in colluvium originated from previous rock slides and 3 debris flows. Large boulders pose a direct threat to life and key infrastructure, amplifying landslide 4 and flood hazards, as they move from the slopes to the river network. Despite the hazard they pose, 5 boulders have not been directly targeted as a mean to detect landslide movement or used in dedicated 6 early warning systems. We use an innovative monitoring system to observe boulder movement 7 occurring in different geomorphological settings, before reaching the river system. Our study focuses 8 on an area in the upper Bhote Koshi catchment northeast of Kathmandu, where the Araniko highway 9 is subjected to periodic landsliding and floods during the monsoons and was heavily affected by 10 coseismic landslides during the 2015 Gorkha earthquake. In the area, damage by boulders to 11 properties, roads and other key infrastructure, such as hydropower plants, is observed every year. We 12 embedded trackers in 23 boulders spread between a landslide body and two debris flow channels, 13 before the monsoon season of 2019. The trackers, equipped with accelerometers, can detect small 14 angular changes in boulders orientation and large forces acting on them. The data can be transmitted 15 in real time, via a long-range wide area network (LoRaWAN®) gateway to a server. Nine of the tagged 16 boulders registered patterns in the accelerometer data compatible with downslope movements. Of 17 these, six lying within the landslide body show small angular changes, indicating a reactivation during 18 the rainfall period and a movement of the landslide mass. Three boulders, located in a debris flow 19 channel, show sharp changes in orientation, likely corresponding to larger free movements and 20


Introduction
proportions of large grain sizes within a landslide mass can also significantly influence its destructive 30 power and affect recovery operations. Large boulders can instantaneously destroy properties, 31 infrastructure and, critically, they can block lifelines for considerable periods of time, as they are the 32 most difficult component of a deposit to remove (e.g. Serna and Panzar, 2018). Boulders can lie on 33 hillslopes for a long time (e.g. Collins and Jibson, 2015), before being remobilised as a consequence of 34 trigger events, such as intense rainfall and earthquakes, which may lead to hazard cascade chains 35 involving boulder transport. In time, boulders have the potential to move from hillslopes and to enter 36 debris flow channels and eventually rivers, posing a hazard along the way. Among the far-reaching 37 effects of boulder movements, damage to hydropower dams can have significant knock-on effects on 38 local economies (e.g. Reynolds, 2018a,b,c). 39 The direct and accurate monitoring of boulder movement, also in relation to environmental variables, 40 is essential in order to achieve a better understanding of the implications of their presence on 41 hillslopes in active landscapes, the dynamics of their remobilisation and their eventual entrainment in 42 river systems. In this context, boulder tracking and real-time monitoring represents an important step 43 forward towards increased resilience in hazard prone areas and it could be performed in different 44 geomorphological settings, ranging from landslide bodies, to loose slope deposits, to debris flow 45 channels and rivers, depending on the specific needs and aims. The ability to produce alerts for either 46 hazardous boulder movements, or to use the movement of boulders to identify hazardous reactivations of existing large instabilities, requires the careful choice of monitoring techniques that 48 work in difficult and different environments, preferably wireless and that can reliably send information 49 in real time. Whilst various early warning systems have been experimented with and put in place for 50 landslides and debris flows, no early warning system has been used to detect and monitor large 51 boulders, thus improving resilience with respect to the additional hazards they pose. 52 Several techniques exist to monitor landslide movements, used also in the context of real time 53 extraction of displacements. For example, early warning systems have been based on traditional 54 techniques such as topographic benchmarks, or extensometers, often in combination with more 55 advanced techniques such as ground based radar interferometry (GB-InSAR) (e.g. Intrieri et al., 2012;56 Loew et al., 2017). Geodetic techniques based on GPS or total stations are also widely used and 57 documented to remotely monitor surface displacements of active landslides (e.g. Glueer et al., 2019). 58 On one hand, traditional techniques tend to be cheaper but they only allow the retrieval of point-like 59 information and they can pose challenges for installation. On the other hand, advanced techniques 60 such as GB-InSAR allow for more continuous coverage but involve much higher costs related to both 61 equipment and data processing and cannot easily deliver information in real time, even if recent 62 research has shown the use of radar techniques to deliver real-time data aimed at rockfall hazard 63 mitigation (Wahlen et al., 2020). Wireless technologies are desirable, due to unfavourable terrain 64 conditions in which landslide monitoring is often needed. In this respect, passive radio-frequency 65 (RFID) techniques have recently been used to monitor landslide displacements, and they have been 66 shown to be inexpensive and versatile (Le Breton et al., 2019). Although this type of technique has not 67 yet been used in early warning systems, it is contended that the adaptability of such technology could 68 be developed in this context. The main advantage is their low cost, their wireless nature and also the 69 ability of the sensors to work in the presence of adverse environmental factors, that would impair 70 other techniques such as GPS and total stations (e.g. fog, snow, dense vegetation). However, passive 71 event has highlighted the need for improved ways of understanding the interactions of cascading 200 hydro-geomorphic processes and to improve measures aimed at increasing resilience (Reynolds,201 2018a,c). The availability of loose material on hillslopes, the monsoonal climate and the GLOFs hazard 202 in the area, enhance the possibility of material containing large grain sizes to reach the river network 203 via hillslope movements, and eventually be remobilised by exceptionally large floods. Huber et al. 204 (2020) highlight that very large boulders (around 10 m in diameter) present today in the Bhote Koshi 205 river have likely been transported by large GLOFs events, supporting the idea that it is unlikely that 206 monsoon generated floods may have the energy threshold required to remobilise very large grain 207 sizes (Cook et al., 2018). 208 Landslides and debris flows can occur also as a consequence of heavy and persistent rainfall during 209 the monsoon. Every year the area receives up to 4100 mm of rainfall between June and September 210 (Tanoli et al., 2017 Rai et al., 2017;Upreti, 1999;Gansser, 1964). 228

Economic assets in the study area -increased vulnerability 229
Our study sites are located along the Araniko Highway, a major route that connects Kathmandu to 230 Kodari and then links Nepal to China. This main road was significantly affected by earthquake induced 231 landslides in 2015, but is also subjected to landslides every year during the monsoon season (e.g. 232 Whitworth et al., 2020). The area is of strategic importance for Nepal due to the high concentration 233 of hydropower projects, either already in operation or under construction (Khanal et al., 2015). 234 Moreover, the Araniko Highway is a key trade and transport link (Liu et al., 2020)

Selected sites 241
The study site is located at the northern edge of an inferred deep seated gravitational slope 242 deformation around 1.5 km wide that stretches from Hindi in the north to just upstream of Chakhu to 243 the south (Reynolds, 2018c). A secondary landslide body on the northwest-facing valley flank directly 244 impinging the settlement of Hindi, and two debris flow channels were chosen as tagging sites (Fig. 1). 245 The most active debris flow channel of the two marks the northeastern boundary of the landslide, 246 whilst the other channel, which appears to be less active, is located 360 m to the northeast, directly 247 upstream of the densest part of the settlement of Hindi. Both channels intersect the Araniko highway 248 and cross the settlement before merging with the Bhote Koshi. The landslide is a soil slide covering an 249 area of approximately 0.03 km 2 . Colluvium material likely deposited from previous landslides is visible 250 at the headscarp and in the terraces along the southwestern flank, with the presence of large boulders 251 of diameter > 2 m. Large boulders are also observed scattered over the landslide body. The scarp 252 suggests a depth of the landslide of at least 2 m, and large, fresh cracks were observed in the crown 253 area in October 2019, indicating activity during the previous monsoon season. 254 Area Network) specification, provided with external GPS and LoRa antennae and measuring 23 mm by 258 13 mm (Fig. 2B), were used as nodes in the system. The sensors are equipped with an accelerometer 259 configured to sample at 2 Hz, as well as a GPS module. In the absence of movement, the devices are 260 programmed to record and transmit one single location (GPS data only) per day at a fixed time. When 261 movement is detected by the accelerometer, so that tilt or acceleration exceed defined thresholds, 262 collection of GPS and accelerometer data is activated. Different thresholds can be applied for a 263 detected angular variation in degrees or for a linear acceleration in g -3 . The values assigned for this 264 study can be found in section 3.3. The sensors, which were developed by Movetech Telemetry and 265

Methodology
Miromico, transmit the acquired data to a LoRaWAN® gateway on the 868 MHz band wirelessly and 266 in real time. A Multitech IP67 LoRaWAN® gateway, sends the payloads received from the sensors to a 267 Loriot LoRaWAN® network server through the local GSM network using an agnostic SIM card ( Fig was filled with epoxy resin, sealing the cavity, thus protecting the device from tampering and from the 275 elements (water and humidity), whilst allowing for unaffected connectivity to the gateway via LoRa. 276 To ease the drilling process but also to allow the epoxy to stay in the cavity before being completely 277 cured, the holes were drilled at an almost vertical angle (with respect to the global inertial frame), so 278 roughly from top down. This allowed for the emplacement of the devices flat against the battery inside 279 the cavity, with z axis near horizontal (global inertial frame), where x and y are oriented as the two 280 longest sides of the device. There is some variability around the deviation from global horizontal of 281 the z axes of all our devices, but in general terms the position of the device would follow such setup. 282 The orientation of the z axis with respect to the cardinal points was not recorded. 283 The position of the gateway, located in the opposite side of the valley at a distance of about 700 m 284 from the furthest sensor, at 1330 m a.s.l. and roughly 60 m above the valley bottom was chosen to be 285 within reach of the GSM network and have direct line of sight with the sensors ( Fig. 1 and 2E). Due to 286 unreliable mains power supply, a 4-panel solar system was developed for this purpose. The initial set-287 up did not allow for continuous power to the gateway and led to instability in the system with frequent 288 offline times during the 2019 monsoon season. However, the system has been improved and it will 289 guarantee continuous power to the gateway for successive acquisition seasons. The panels currently 290 charge two 12 V, 110 AH batteries that then provide continuous power to the gateway through a POE 291 (power over ethernet) supply. The solar system is composed by parts that can be sourced locally, at 292 relatively low cost and that can be transported to sites without road access, such as the site chosen in 293 this study. The nature of the local GSM network, relying on one individual antenna in the area at the 294 time of this study, has also led to frequent GSM connection failures which prevented the gateway 295 from communicating with the server. The devices deployed in the 2019 season were programmed to 296 not store the data, but to send it immediately, causing the data transmitted during gateway offline 297 time to be lost. 298

Choice of tracked boulders 299
The tagging sites were selected with the aim of covering different geomorphological settings whilst 300 retaining visibility to the gateway. The boulders identified for tagging are spread over three sites, two 301 debris flow channels and a landslide body (Fig. 1). The boulders cover a range of sizes and geologies, 302 though the geology in this context is not expected to play a significant role in affecting the connectivity if the soil were to be eroded during intense rainfall events. 320

Sensors settings 321
The sensors were programmed to send a routine message every 24 hours, in which only the GPS 322 position is sent. In between regular fixes the sensors sleep and do not send any data unless movement 323 occurs, as explained in the following text. As mentioned in section 3.1, the sensors can also acquire 324 and send data in association with an accelerometer event for which activation thresholds can be set 325 for impact forces and for angular variations. The sensors can be programmed following two main 326 modes: 1) the accelerometer data is averaged over a window of time (over a number of recordings), 327 we call this mode "average" settings (AVG in Appendix 2) and 2) the absolute value of the maximum 328 acceleration occurring in a time interval can be recorded, and we call this mode "maximum" settings 329 (MAX in Appendix 2). In the first case, the values of the three axes are normalised to g force (where 1 330 = 1 g) and the measurements essentially represent the static angle of tilt or inclination, thus the 331 projection of the acceleration of gravity, g, on the three axes, ranging between 0 (for an axis oriented 332 horizontally with respect to the global inertial frame) and ± 1 (for an axis oriented vertically with 333 respect to the global inertial frame). In the second case, the absolute maximum value can be recorded 334 and this can exceed 1 g and can be set to be as high as 2, 4, 8 or 16 g. The measurement resolution 335 changes according to the chosen detectable maximum, so that a scale capped at 2 g has a resolution 336 of 0.016 g, whilst a scale capped at 16 g has a resolution of 0.184 g (Appendix 3). 337 When considering only an individual axis, the variation between two static accelerometer 338 measurements would correspond to an angular change as shown in Eq. (1): 339 where is the angular variation on a given axis and is the difference between normalised successive 343 accelerometer values recorded on the same axis in g. Eq. (1) describes the relationship between 344 accelerometer output on a given axis and its tilt: for trigonometry, the projection of the gravity vector 345 on an axis produces an acceleration that is equal to the sine of the angle between that axis and a plane 346 perpendicular to gravity. According to Eq. (1), if the scale is capped at 2 g, for = 0.016 g the 347 corresponding angular variation is of approximately 0.9° if the axis is vertical (with respect to global 348 inertial frame), but approximately 5.5° if the axis approaches horizontal. Similarly, if the scale is capped 349 at 16 g, a value of = 0.184 g corresponds to an angular variation of about 10° when the axis is near-350 vertical, but this increases to as high as approximately 21° when the axis approaches the horizontal 351 The boulders expected to move as a whole with the soil in which they are embedded, and that are 353 more likely to experience small and gradual angular variations as the surrounding material gently 354 slides, were programmed with the average settings. We chose to cap accelerometer data for average 355 settings at 2 g (highest resolution), as high impact forces were not expected, and we assigned 356 thresholds for activation on accelerometer events of approximately 0.4 g and 5° for impact forces and 357 angular changes respectively. The sensors in the two debris flow channels and some of those only 358 partly embedded within the landslide were programmed to record high impact forces using the 359 maximum settings (Appendix 2). In this case, the scale was capped at the maximum detectable force 360 of 16 g (lowest resolution) and the impact and angular thresholds were set at approximately 4 g and 361 5° respectively. This angular threshold yielded noisier data with respect to the sensors programmed 362 with the average settings type, because of the direct consequence of a drastic reduction in 363 measurement resolution in the sensors programmed with the maximum settings type (Appendix 3), 364 for which the scale was capped at 16 g. Natural measurement variability and errors associated with 365 the sensors led to spurious data, given the relatively small angular threshold assigned for the highest 366 detectable maximum of 16 g. In other words, given that the step of accelerometer measurement is as 367 high as 0.184 g, a spurious angular variation of more than 5° is often detected even when the boulder 368 is stable, due to intrinsic measurement variability (up to 2 bits). Due to the fact that an angular 369 threshold lower than the scale resolution was imposed, we observed many extra acquisitions triggered 370 by small variability in accelerometer measurements around a stable value, rather than by true 371 movement. 372 In order to reduce the noise in the data due to these fluctuations, a three-stages smoothing is applied 373 to the raw data. First, a moving window covering 5 successive data points is used. The median value 374 of the 5 data points is assigned to all points in the window that lie within ± 0.184 g of the data point 375 immediately before the window. If any of the values lie outside the ± 0.184 g threshold, then the raw 376 data points are left unchanged. In the second stage, peaks of one data point are removed (i.e. one 377 point above or below two points with the same value), this is because if a high impact force is imparted 378 to a boulder, the position of the boulder is expected to change. This would mean that a high value 379 would likely be followed by a change in the static angle of tilt of the three axes. Therefore, it is 380 unrealistic to have a peak value followed by a value equal to that observed before the peak, 381 particularly when sampling at 2 Hz. This would imply that a boulder undergoes acceleration in one 382 direction, moves and comes to a halt in the same orientation as before the movement. In the third 383 and final stage, another moving window of 5 consecutive data points searches for values that lie within 384 The sensors are equipped with a GPS module, which is currently also used to retrieve the date and 405 time of the data acquisition, whilst the data transmission has another timestamp related to the arrival 406 of the data string to the server. The accelerometer readout in the current version of the software is 407 tied to a GPS acquisition, this means that although the accelerometer is activated as soon as 408 movement is detected, the recording of the acquisition is obtained only when the GPS has successfully 409 retrieved the position. An acquisition of accelerometer data with no GPS position can be obtained and 410 transmitted (in which case it would only be associated with a server timestamp indicating time of 411 arrival at the server), but only after the GPS has attempted to retrieve the position and failed. The 412 timeout for the GPS search has been set to 120 seconds. This is because due to the local topographic 413 setting and the high valley flanks, the availability of enough satellites at any given time may be low. to red symbols in Fig. 4). Of these, six lie within the landslide body and were programmed with the 454 average settings in order to detect small angular changes (Fig. 5). The remaining three were located 455 within the southern debris flow channel and were programmed with the maximum settings, to capture 456 large (> 1 g) impacts (Fig. 6). 457 In terms of boulder sizes, boulders that appeared to have moved within the landslide have b-axes 458 ranging from 0.4 to 2.75 m, whilst those that moved in the southern channel have b-axes comprised 459 between 0.4 and 0.5 m (Appendix 1), thus covering a much smaller range. 460 The 4 boulders within the landslide that do not show evidence of movement (white circles in Fig. 4), 461 were fitted with sensors programmed with the maximum settings (Appendix 2), due to the fact that 462 they are partly embedded in the landslide and had potential to become detached from the landslide 463 body, and thus given the lower accuracy and coarser scale they could not have detected small, gradual 464 movements even if they had been subjected to them. ). The grey curves are raw data and the yellow, orange and red curves are the data after noise was 475 removed. The data is actual data recorded by the accelerometers, referred to a common zero for 476 visualisation purposes, as explained in section 3.3 (hence all raw data curves begin at 0, and the 477 smoothed curves around zero, due to the smoothing). A sketch of the possible type of movement 478 related to gentle tilting of the boulder within the soil mass, is shown in panels A and B in Fig. 5 and  479 does not represent any true movement of any of the tagged boulders. The data shows that all sensors 480 that detected movement were appropriately charged throughout the season (blue curves in graphs).  Fig. 4). 507 Displacements roughly up to 2 m in the image plane are detected in the lower and mid-slope parts of 508 the moving area ( Fig. 5H and 7A) between the end of August and the beginning of September, with 509 upper parts showing displacements of around 1 m. The movement observed in the accelerometer 510 data of B# A226 and B# 9A41 (Fig. 5F-G) corresponds to the periods in which higher displacement 511 magnitudes are inferred from the images. Fig. 4 and Fig. 7B also show that boulders B# 5B6A, B# 33EB 512 and B# 9A41 are located in areas surrounded by displacements as seen by the TLS data (yellow hatched 513 areas in Fig. 4). Moreover, two boulders within the upper part of the landslide were not found in the 514 field campaign carried out in October 2019 (B# 33EB and B# 625C), likely due to fresh accumulation of 515 material from the scarp. Indeed, TLS scan data show cumulative displacements of up to 1 m over large 516 areas between April and October 2019 (Fig. 7). 517  Fig. 6A, B, C contain 520 the same accelerometer information as explained in section 4.1. The difference in the scale of the 521 accelerometer output with respect to Fig. 5 is explained by the different settings. These boulders were 522 programmed to retrieve accelerations higher than 1 g (as opposed to normalised values) and forces 523 up to 16 g. The raw data (grey curves) show frequent oscillations often within ± 0.184 g around a value 524 (corresponding to one step in the accelerometer scale, or one bit) and occasionally up to ± 0.372 g 525 (two steps in the scale, two bits), associated with measurement variability and the coarse scale used 526 (see section 3.3). 527

Rapid orientation changes of boulders in the southern debris flow channel 518
As an example, in the graph for B# 4C02, we observe a change from the initial orientation of the 528 accelerometer within the boulder equivalent to 1000 mg in y and around 700 mg in x and z. This is 529 compatible with a change between the initial orientation (1) and orientation 2, attained by the boulder 530 by 4 June 2019, as visualised in Fig. 6B. The current settings have not captured how the boulder 531 transitioned between position 1 and position 2, likely due to the very short time interval during which 532 the change is expected to have happened. The GPS acquisition is likely to have taken longer than the 533 movement that triggered the recording and delayed the accelerometer acquisition. This applies to the 534 other two boulders shown in Fig. 6. We do not observe forces > 1 g for any of the sensors programmed 535 with the maximum settings, despite the ability of the sensors to detect up to 16 g. This is consistent 536 with a lack of debris flow activity recorded by cameras or seismometers, the more prolonged activity 537 of which would have generated sustained boulder movement, beyond the time needed for GPS 538 acquisition as explained below. 539 The vertical green bars in the graphs of B# 57B9 and B# FB58 (Fig. 6C and E)  and server time stamp are the same (data sent in real time). Thus, the second movement of B# FB58 580 is likely to have occurred between these two times, even if the data acquired after the gateway was 581 online again has been sent in real time on 25 August. The camera images show that movements on 582 the right bank of the channel occur between 22 and 24 August. The scan data also shows important 583 displacements in the channel right bank (Fig. 8C). Moreover, 5 boulders in the channel (or on the bank) 584 were not found in October 2019 at their original location. Two of these are boulders that appear to 585 have moved in the smart sensors' data and the other three may have been covered by deposition of 586 loose material. 587 No boulder movement was recorded for the northern channel, and field observations in October 2019 588 revealed no signs of recent activity in the channel, which was completely overgrown with vegetation. 589

GPS module limitation 590
The GPS had an overall poor performance across all the sensors during the data acquisition season. The GPS data acquired is unrealistic not only for the magnitude of the position differences of the same 596 boulder, but also because the direction is often inverted in time, which is not compatible with possible 597 boulder movement. However, the poor performance of the GPS for the purpose of boulder tracking 598 has only limited impact on the ability to detect movement or orientation changes using the 599 accelerometer, as outlined in the previous sections. 600 The movements observed for the boulders scattered on the landslide body and embedded within the 611 material can be described as small angular variations that occurred gradually during the season. Visual 612 recognition of such movements in the field or in the camera images and scan data would be unfeasible 613 for individual boulders because they correspond only to small tilt that is difficult to detect with such 614 methods. However, there are elements that support the fact that the data acquired by the 615 accelerometers is real and caused by gradual tilting. The images acquired by the camera show 616 important sliding of the landslide up to 2 m in August-September (see section 4.1 and Fig. 7A), when 617 the boulders located around the southwestern flank and in the lower part of the landslide show higher 618 magnitude of the angular variations with respect to other boulders (Fig. 5F, G). The fact that the onset 619 of movement observed in six boulders in the landslide is not random but appear to follow a spatial 620 and temporal pattern also supports the idea of a landslide reactivation that causes smaller movements 621 around the headscarp and nearer the channel to occur earlier. The headscarp activity may not only be 622 related to the movement of the entire mass, but also to small collapses of the colluvium material in 623 the steep exposure. This may have led to small movements already from the onset of the monsoon. 624

Discussion
Movements in this area are supported by data obtained with the TLS that indicate that displacements 625 in the line of sight of up to 1 m occurred at or just below the headscarp during the season (Fig. 7B). 626 Moreover, two boulders in this area were not found in October 2019, most likely because they have 627 been covered by collapses of loose material from the headscarp. The area near the northeastern flank 628 may have experienced an increase in pore pressures due to earlier saturation of the soil here than in 629 the area at the opposite flank, also related to a more rapid increase of the ground water table nearer 630 the channel driven by topography. We also observe that the magnitude of movements of boulders 631 closer to the southwestern flank and in the lower slope is higher than elsewhere; this is well supported 632 by observations obtained through the field camera. 633 Four partly embedded boulders in the landslide (Appendix 2) were programmed with the maximum 634 settings and showed no movement (Fig. 4). The reason to choose this setting type for these boulders 635 is that the nature of their position (PE) may have led to larger and faster downslope movements if 636 they had become dislodged. Given the lower resolution of the data obtainable from the maximum 637 settings, it is possible that nothing is observed for these boulders even if they moved consistently with 638 the landslide body and experienced slow and gradual tilting of a few degrees. In other words, it is 639 possible that such boulders also moved but that the nature of the movements may have been too 640 subtle to be captured with the settings applied. It is also possible that these boulders found 641 themselves outside of the active sectors of the landslide, although this seems less likely given the 642 observations obtained in the field and also from camera images and scan data. Although camera 643 images, scan data and accelerometer data are characterised by different time resolutions, the 644 movements observed in both landslide and channel in the images and the amount of erosion and 645 deposition observed in the scan data indicate that the boulders tagged were likely involved in such 646 movements, and thus there is increased confidence in the fact that the accelerometer data indeed 647 indicate real movement of the boulders. 648 Another element that supports the fact that the recorded accelerometer data is associated with real 649 boulder movement is related to boulder size. Appendix 1 shows boulder sizes for boulders with and 650 without movement in the three different tagging sites. For boulders within the landslide body, a size 651 control on movement was not anticipated. This is because boulders were expected to move as a whole 652 with the landslide mass and thus their potential to be transported would be independent from their 653 size. On the contrary, in the channel, and particularly for boulders lying in the channel bed, a size 654 control on movement is expected, because the size of boulders that could be mobilised by a flow 655 depends on the flow intensity (Clarke, 1996). Therefore, a flow with low intensity could not be 656 expected to mobilise the largest boulders tagged. The observations indicate that boulders that show 657 movements in the landslide are characterised by a much higher range of b-axes than those in the 658 channel (Appendix 1). 659 For boulders programmed with the maximum settings, we observed noisier accelerometer data than 660 for those programmed with the average settings. What controls this behaviour is not the fact that the 661 sensors were programmed to detect the maximum force or the static tilt respectively, but rather the 662 scale that was chosen and associated with the two settings types combined with the choice of angular 663 threshold to trigger acquisitions. As mentioned before, 16 g and 2 g were chosen as values to cap the 664 scale in the maximum and average settings respectively. 665 When a sensor is programmed to be capable of capturing forces impacting a boulder as high as 16 g, 666 the resolution currently available for the accelerometer's reading is of 0.184 g. Although this is a 667 relatively small value with respect to 16 g, this corresponds to an angular variation of 10.7°. Moreover, 668 we observe that measurement variability is often 1 bit, but occasionally 2 bits, the latter corresponding 669 The acquisition of a GPS position is also what causes the largest battery expenditure in the sensors 689 and it is therefore detrimental for long-term data acquisition on boulder movement. The high 690 positional errors and the important battery expenditure make the current GPS module not fit for the 691 purpose of tracking boulders in rugged terrains. 692 As mentioned above, it is possible to retrieve data strings from the sensors without a GPS timestamp. 693 So, even if a GPS position, date and time cannot be acquired, the accelerometer data can be recorded 694 and transmitted anyway, with the server timestamp. In this sense, the fact that the accelerometer was 695 tied to the GPS during the 2019 acquisition season, so that the accelerometer data could be recorded 696 only once the GPS acquisition has been attempted and failed, did not invalidate completely the data 697 output. 698 However, there are also important limitations related to this. As the time for the GPS acquisition 699 attempt was set to 120 seconds, the sensor measures the acceleration already during this time, but it 700 does not record it nor transmit it until the GPS position is either acquired or fails. In the case of fast 701 movements, or relatively large impacts caused by the sudden movements of boulders within the flow, 702 120 seconds (this would often be even more, in case a GPS acquisition is being obtained) may be 703 enough time for the movement to begin and stop. This may explain why, although the boulders in the 704 channel were programmed to detect high forces, they never show accelerometer values higher than 705 1 g (either negative or positive). In essence, these sensors have also only recorded the static tilt and 706 different orientations acquired by the boulders in time (within seconds of movement occurrence), but 707 not the actual movement as it unfolded. For instance, the position change of B# 4C02, B# 57B9 (second 708 event, i.e. event that causes transition from position 2 and 3) and B# FB58 (first event, i.e. event that 709 causes transition from position 1 and 2) were received in real time. This means that as soon as the 710 data string indicating a different orientation with respect to the previous data string was acquired, it 711 was also sent. In this type of situation, the GPS timestamp is the same as the server timestamp, but 712 there is no recording of the movement as it unfolded. The event of B# 4C02 points to the fact that the 713 GPS delayed the acquisition of the accelerometer data, because the gateway was online during the 714 time in which the orientation change must have occurred. Given that there is no evidence of large 715 debris flows during the 2019 monsoon season, B# 4C02 may just be one example of minor boulder 716 movement that started and stopped within the 120 seconds time interval. This may be improved in 717 successive acquisition seasons, since development has been made in order to separate the GPS from 718 the accelerometer acquisitions. The next batch of devices that will be deployed in the network will 719 thus be able to capture faster rotation already from the start of the movement. 720 The picture may be complicated even further by the fact that occasionally the gateway experienced 721 some offline time, due either to the battery not being recharged properly or to GSM connection loss. 722 This is the case of B# 57B9 (second event) and B# FB58 (first event), in which we observe that the data 723 string indicating an orientation change is sent in real time, but follows a gap in the gateway 724 connectivity. In this case, the movement may have occurred at any point during the offline period of 725 the gateway, then the first acquisition since the gateway became once again online is sent in real time. 726 However, a new solar system is now in place and will prevent future power issues during future 727 acquisition seasons. Finally, the accelerometer sampling acquisitions that could be reached in the 2019 728 campaign was 2 Hz. While this is acceptable to detect gradual angular variations that occur slowly over 729 a prolonged period and allowed us to identify periods of acceleration of the rotations, it is too low if 730 the aim is that of capturing a fast movement in the channel. For this reason, the capability of our 731 devices has now been increased to record data up to 400 Hz. 732

Advantages and limitations of this technology 733
The LoRaWAN® smart active sensors developed in this study for the purpose of identifying boulder 734 movements has already shed light on its potential advantages and its limitations. The technology used 735 is independent of weather conditions. The communication between the tags and the gateway is not 736 hampered by adverse weather conditions and movements were observed during overcast and rainy 737 days. This is of course true if the gateway is powered with batteries of sufficient capacity to withstand 738 days with insufficient sunlight, which may occur during the monsoon season. Although a good visibility 739 of the sensors from the gateway increases connectivity between the nodes and the gateway, the long-740 range nature of the system allows for a network that extends over a relatively large area. In our case, 741 we were able to obtain data from boulders located at up to 800 m from the gateway, covering an area 742 of about 0.25 km 2 , this likely not being the upper limit of the achievable range. This is especially 743 advantageous for a number of reasons. Different geomorphic features can be monitored with the 744 same gateway, in our case including a landslide and two debris flow channels. Moreover An important characteristic of the devices used in this study as opposed with other techniques is that 755 they are active and can easily be assigned thresholds (e.g. acceleration or tilt) that can be used in an 756 early warning system context. Moreover, the devices can be embedded directly inside boulders, 757 without the need for additional supports that may 1) make the devices more visible/exposed and thus 758 more subjected to intentional tampering or animal damage, 2) there is no additional movement to be 759 accounted for (e.g. tilting of supporting poles). The technology is also relatively low cost and has the 760 potential to become competitive and cost-effective in the future. The most expensive component is 761 the gateway (around 1000 USD), whilst the devices are around 200 USD each. The ability to retrieve 762 the tags after battery consumption has already been investigated and will be implemented in 763 successive acquisition seasons, will allow for a durable, cost-effective network. This may make this 764 technology more affordable than other more expensive techniques such as GB-InSAR, GPS or total 765 stations and can allow dense networks. 766 The main drawback encountered in this study is the poor performance of the GPS module, which made accelerometer data, has hampered in some cases the ability to obtain the full sequence of 773 accelerations experienced by the boulders. This issue will however be resolved in the next acquisition 774 season, since further development has allowed us to make the accelerometer independent of GPS 775 acquisitions. Work is also planned to write the firmware to enable the gyroscope and magnetometer 776 on the device, which will give more detail of boulder dynamics such as rotations. Finally, the 777 connectivity of the gateway to the server (during offline periods) has prevented some of the time the 778 ability to receive the movement signal in real time. This problem has now been resolved, with a more 779 stable solar system currently powering the gateway, thus future acquisition seasons should benefit of 780 higher robustness and less connectivity loss. reactivation of the landslide is confirmed by both timelapse cameras and TLS data. Also, the 790 movements show staggered onset, so that the boulders nearer the scarp or the lower boundary, near 791 the channel, began to move earlier in the season than other boulders. In the channel, only three 792 boulders show data likely corresponding to sharp, sudden movements and rotations that occurred in 793 response to intense or persistent rainfall. The sizes of the boulders that moved in the channel are 794 towards the smallest end of the boulders tagged in the channel, reflecting the fact that no large debris 795 flows were observed in the channel during the 2019 monsoon season. 796 Though with some limitations, the technology has proven able to detect boulder movements with this 797 type of device, for the first time in a field setup as opposed to a laboratory setup. In the optimal 798 conditions of all the component of the network operating properly, the ability to capture the onset of 799 movement in real-time is an important premise to the use of this technology in early warning systems 800 of slope movements that involve the presence of hazardous boulders. This pilot study also hints at the 801 potential of these devices to further understanding of landslide dynamics, for example the timing of 802 movement in response to rainfall and the spatial sequencing of movement across a landslide. The 803 most important challenge that we believe has prevented the recording of the complete movement for 804 the boulders in the channel is related to the current requirement for a GPS position to be acquired for 805 the accelerometer data to be recorded and transmitted. Furthermore, the poor GPS performance 806 currently precludes the measurement of displacements. However, the sensors are already equipped 807 with a 9-axis IMU comprising an accelerometer, a gyroscope and a magnetometer, that have not been 808 ready for the field tests in Nepal, that might allow the retrieval of more information on movement, 809 when combined with field observations and optical images. 810 Future work will involve the tagging of more boulders at the same sites of the current network to 811 improve the accelerometer sampling frequency, the now improved the stability of the network 812 connectivity, more suitable programming settings and the ability to retrieve and reuse the tags. In the 813 next batch of devices, we will be able to activate the accelerometer and record movement data 814 independently of the GPS acquisition. This is expected to significantly speed up data acquisition and 815 transmission to the server, which will be a step forward in view of using this technology for early 816 warnings. Moreover, this will also allow us to capture the whole accelerations sequence associated 817