A decade of detailed observations (2008–2018) in steep bedrock permafrost at Matterhorn Hörnligrat (Zermatt, CH)

The PermaSense project is an ongoing interdisciplinary effort between geo-science and engineering disciplines and started in 2006 with the goals to realize observations that previously have not been possible. Specifically, the aims are to obtain measurements in unprecedented quantity and quality based on technological advances. This paper describes a unique ten+ year data record obtained from in-situ measurements in steep bedrock permafrost in an Alpine environment on the Matterhorn Hörnligrat, Zermatt Switzerland at 3500 m a.s.l.. Through the utilization of state-of-the-art wireless sensor technology it was 5 possible to obtain more data of higher quality, make this data available in near real-time and tightly monitor and control the running experiments. This data set (DOI: doi.pangaea.de/10.1594/PANGAEA.897640, Weber et al., 2019) constitutes the longest, densest and most diverse data record in the history of mountain permafrost research worldwide with 17 different sensor types used at 29 distinct sensor locations consisting of over 114.5 million data points captured over a period of ten+ years. By documenting and sharing this data in this form we contribute to making our past research reproducible and facilitate 10 future research based on this data e.g. in the area of analysis methodology, comparative studies, assessment of change in the environment, natural hazard warning and the development of process models. Finally, the cross-validation of four different data types clearly indicates the dominance of thawing-related kinematics.


Introduction
The behavior of frozen rock masses in steep bedrock permafrost rock slopes is a dominant factor influencing slope stability 15 when permafrost warms or thaws (Fischer et al., 2006;Ravanel and Deline, 2014). Ongoing degradation of mountain permafrost coincides with observations of increasing rockfall activity (Ravanel and Deline, 2011;Huggel et al., 2012;Gobiet et al., 2014) potentially triggering large scale hazard events via complex process chains (Huggel et al., 2005;Westoby et al., 2014;Haeberli et al., 2017). While the long-term trend of rising permafrost temperatures can clearly be observed at a global scale (Biskaborn et al., 2019) warming trends of mountain permafrost are more diverse in their behavior . For example it has been recently observed that the generally warming trend can be temporarily interrupted depending on the amount and temporal extent of the snow cover (Noetzli et al., 2019) which is especially variable in mountainous terrain.
The Matterhorn Hörnligrat field site located in Zermatt, Switzerland at 3500 m a.s.l. is a unique situation for steep bedrock permafrost research as it is located on a ridge and not on a mountain top or in a large rock face where permafrost boreholes would typically be placed (Luethi and Phillips, 2016). A comprehensive multi-sensor setup has enabled research on surface 15 processes and kinematics in steep bedrock permafrost in the context of environmental forcing (ambient meteorological conditions, snow cover, heat flux) since 2006. Situated in a unique and iconic setting, the Matterhorn Hörnligrat field site now provides over a decade of mountain permafrost data: the longest, densest and most diverse data record with respect to permanent monitoring of mountain permafrost at high elevation worldwide. Apart from duration and location, this data set is novel with respect to the diversity of the instruments used (17 different sensor types are contained in this paper), the density of the 20 measurements both spatially (sensors are installed at 29 distinct sensor locations each containing one or more sensor types (see Table 2) and temporally (sampling rates on the order of per-minute to per-second). The data set presented amounts to 83.8 GB of data in 41'031 files of different formats containing approximately 114.4 × 10 6 data points of primary and aggregated data (see Table 6). To the best of our knowledge, in the entire European Alps only the Aiguille du Midi site (Chamonix, France, 3842 m a.s.l.) (Ravanel and Deline, 2011;Magnin et al., 2015), the permafrost borehole on Jungfraujoch (Grindelwald, 25 Switzerland, 3700 m a.s.l.) (Wegmann, 1997(Wegmann, , 1998Noetzli et al., 2019), the geothermal profiles on Stockhorn (Gruber et al., 2004c) and two simple ground surface temperature sensors located on the summit of Matterhorn (4478 m a.s.l.) are located in comparable or at higher altitude and are being operated in a long-term monitoring mode albeit the data records are shorter and offer less diversity with respect to the measurements. Other study sites at very high altitude exist, e.g. Grandes Jorasses (Chamonix, France, 4208 m a.s.l.) (Faillettaz et al., 2016) but have only been operated for a short period and in campaign mode. 30 Outside of the European Alps, mountain permafrost data is very sparse (even in the Himalaya, Gruber et al., 2017) and in cases where ground-based measurements exist they are likely limited to a single sensor type only (Zhao et al., 2010;Popescu, 2018;Gruber et al., 2015).
This manuscript documents the complete raw data at full sampling rates of the instruments used (primary data set, see Section 3) for the most significant sensor channels/types deployed (as outlined in Section 4) as well as a selection of derived 35 data products (secondary data set). The derived data products are downsampled and cleaned time series of the weather station, ground temperature, electrical resistivity of rock, fracture displacement and inclinometer data as well as GNSS daily positions computed using double differencing techniques. In order to be able to fully understand and leverage the high-fidelity sensor and to allow full transparency and reproducibility a technology excerpt as well as the procedures for compiling and validating the primary and secondary data sets are presented in Section 3 and Section 5 respectively. Using the toolset described in 5 Section 7 these data sets can be recreated and independently updated (living data process). The online data portal at http: //data.permasense.ch (see Figure B1) is discussed in Appendix B. In addition, select examples of the data as well as an overview of the scientific results based on data from this field site are discussed in Section 6.

Matterhorn Hörnligrat field site
The Matterhorn is prominently known due to its archetypical form, the famous climbing route up the Hörnligrat ridge (northeast 10 ridge of Matterhorn) and its dramatic first ascent on July 14, 1865. This first successful alpine conquests on the Matterhorn were actually undertaken by researchers: first ascent led by Edward Whymper, a writer and landscape illustrator on assignment from an English publishing house as well as subsequently the second ascent by John Tyndall, a prominent multidisciplinary scientist of his time both accompanied by local guides and other companions. Nowadays, several hundreds to few thousands of mountaineers climb the Matterhorn via the Hörnligrat every year. 15 In the exceptionally warm summer of 2003 increased rockfall activity was observed in the entire Alps (Gruber et al., 2004a;Ravanel et al., 2017). An increasing interest into the thermal behavior of permafrost in steep topographies in these years (Gruber et al., 2004b, c) lead to a first simplified modelling study based on the Matterhorn (Noetzli et al., 2007). It soon became clear that such work would require substantial evidence from long-term, in-situ measurements to calibrate and validate such models accordingly as no other comparable data set existed. Additionally, the prominent rockfall activity observed motivated further 20 research questions with respect to slope/rock wall stability, natural hazards (mitigation) and the susceptibility of nearby human infrastructure and urban environments to such hazards (Gruber and Haeberli, 2007;Fort et al., 2009;Ravanel et al., 2017Ravanel et al., , 2010 Figure 1a, CH1903+ 617950/92168) uncovering bare ice in the failure plan (see red arrows in Figure 1b and c) (Hasler 25 et al., 2012). Although insignificant on the scale of a mountain the age and size of the Matterhorn as a whole, this particular incident showed significant susceptibility on the human scale to the processes governing such rockfall: as this rockfall event occurred in the middle of the summer climbing season and directly affected the popular climbing route to the summit, it led to the evacuation of 84 climbers by helicopter, the temporary closure of the climbing route and other mitigation measures (Haeberli et al., 2015). Compact ice was observed on the surface of the detachment scar right after the rockfall event suggesting 30 clefts be filled with ice. With respect to the research aspects it is this hazard event, the expectation that further (catastrophic) dynamics would likely follow and the significant ice infill that led to the selection and instrumentation of the first experiments  (Hasler et al., 2008).
Therefore an initial interdisciplinary project between geo-science and engineering was proposed with the initial goals to enable observations that previously have not been possible: the PermaSense project specifically aimed at (i) obtaining in-situ measurements with unprecedented quality and quantity (with respect to both spatial and temporal resolution and duration) but 5 also (ii) to try to leverage then-emerging wireless sensor network technology (Talzi et al., 2007;Hasler et al., 2008) at scale and in a real case study. The Swiss Science Foundation (SNSF) funded National Competence Center on Mobile Information and Communication Systems (NCCR- MICS Aberer et al., 2007) as well as select government funding through the Swiss Federal Office of the Environment (FOEN) supported this initial push that over time development into a comprehensive outdoor infrastructure and mountain lab supporting diverse experiments, long-term monitoring and online data: http://data.permasense.

ch.
The Hörnligrat field site is located at 3500 m on the North-East ridge of Matterhorn covering the area around the detachment zone of the 2003 rockfall event and consists of steep, highly fractured rock slopes with partially debris covered ledges and different expositions, where the expected occurrence of permafrost varies with aspect and relief conditions (see Figure 3 in Weber 5 et al., 2017). Geologically, this field site consists of gneiss and amphibolite of the Dent Blanche nappe (Bucher et al., 2004) and the most dominant fractures are oriented parallel to the ridge and dip nearly vertical (see Figure 2 in Hasler et al., 2012).
Climatically, the region of Zermatt is characterized by a dry and subcontinental climate with high daily/seasonal temperature fluctuations and with mean annual air temperature (MAAT) of 3. 5 • C (1961-1990) and 4.2 • C (1981-1990) (MeteoSwiss, 2019). While reanalysis data with a 1x1 km 2 grid indicate a regional MAAT of −6.7 • C (1961( -1990( , Hiebl et al., 2009) for the 10 field site area, local measurements at the field site show a mean annual air temperature of −3.7 • C (period 2011, Weber et al., 2017. As precipitation mostly falls as snow with occasional infrequent rainfall events in summer, liquid water is mainly supplied to the site by snow melt (Hasler et al., 2012). Winter temperatures down to −27 • C in combination with exposure to strong wind (to over 100 km/h) results in a preferential snow deposition in fractures, on ledges and at other concave microtopographical features. While the northern side contains small ice field within a steep heterogeneous rock face, on the south side 15 snow patches develop during winter in couloirs as well as on rock bands and disappear in spring/summer completely (Hasler et al., 2012).
Surveying and site selection took place in the years 2006/2007 with an initial sensor installation campaign in fall 2007 (Hasler et al., 2008). The technological developments started with data logger prototypes (Talzi et al., 2007) that were used for a first data retrieval campaign during the following winter season. The prototype development and initial experience resulted in a 20 redesign of the wireless sensing platform that was deployed for the first time on July 25, 2008(Beutel et al., 2009). This date also marked the start of the "production" data generation for the PermaSense project and the data contained in this publication. Later technological milestones include the introduction of the GSN data management system, a switch from 3G cellular connectivity to IEEE 802.11a 5 GHz WLAN for long-haul connectivity and the introduction of a middleware software infrastructure for mitigating data loss through back-pressure in summer 2009. On the sensor side extensions took place in 25 2009 with a remote controlled high-resolution visible light camera (Keller et al., 2009b, a) as well as a significant extension of the crackmeters in summer 2010 and the installation of a high-precision survey-grade GNSS receiver at the very end of 2010 (Beutel et al., 2011). A local weather station was added in 2010 and a net total radiometer in 2015. After this first research phase focusing on prototyping and the investigation of surface kinematics with respect to thermal forcing (Hasler et al., 2012;Weber et al., 2017) an additional research avenue was added from 2012 onwards: a first pilot study using acoustic emission and 30 based on similar efforts undertaken at the Jungfraujoch (Amitrano et al., 2012;Weber et al., 2012;Girard et al., 2012Girard et al., , 2013 aimed at characterizing damage evolution inside the solid rock walls in 2012/2013. A larger profiling experiment (Weber et al., 2018c) has been set up to investigate signals emanating from the mountain and possible damage events with different instruments ranging from 1 Hz to 100 kHz as well as additional L1-GPS measurement points starting in 2015/2016. Finally, in an effort to establish a vertical transect of thermal measurements spanning the whole mountain (two ground surface temperature 35 measurement points exist on the summit since 2011, maintained by Agenzia Regionale Protezione Ambiente Valle d'Aosta (ARPA VDA), Italy, permafrost boreholes maintained by PERMOS, SLF/WSL, Switzerland are located on lower elevations at the Hörnlihütte and Hirli) an extension with further ground surface temperature profiles implemented at 4003 m a.s.l. in the vicinity of the Solvay Hut higher up on the ridge has been performed. Despite its remoteness and exposure this field site is actually readily accessible being situated directly on and in the bottom segment of the climbing route with further infrastructure 5 nearby (mountain hut, heliport, transportation facilities, Internet connectivity) and therefore can be accessed even in a day trip from Zurich.
After completing the first ten years since the first experiment went live in July 2008 it's now time to publish a first digest of this data including a thorough documentation in order to (i) preserve this data and (ii) make it available for future research in the broader context. The data presented in this publication constitutes a best-of of the most relevant and descriptive geo-science 10 related data collected. There are further data available in the context of this work, that either (i) have been published elsewhere (Weber et al., 2018a;Meyer et al., 2018), (ii) is not deemed suitable for publication in the context of this publication (either out of scope or to complex or too poor in quality) and (iii) have been collected by related activities in the vicinity of this field site.
The most relevant of these additional data sources are described in brief in Section 4.8 in order to give the reader the relevant pointers in this context. 15

Instrumentation technology and data management
The core instrumentation technology employed at this field site are autonomous, low-power wireless networked sensors (Beutel et al., 2009), frequently also referred to as wireless sensor network or short sensor network. The promise of this novel technology at the time of the conception of this field site in (Hasler et al., 2008 was to allow unobtrusive, large-scale and highly reliable measurements based on a minimum resource footprint without a central point of failure and extensive cabling. 20 Apart from geoscience investigations the first PermaSense project pursued the goal to develop means for long-term, highquality sensing in harsh environments, generating better quality data, with online data access in near real-time (Hasler et al., 2008). Using such technology it would be possible to achieve measurements that previously have not been possible and consequentially to enable new science, answering fundamental questions related to decision making, natural hazard early-warning.
For selected sensors, where the integration as low-power wireless sensor was infeasible or impractical, industry standard com- 25 ponents have been used although they have typically been adapted and integrated with our custom network, data and power management infrastructure based on our sensor network technology. Our experience over the past decade+ shows, that using a WSN is a promising approach with superb data availability and data integrity. The sensor nodes have been running reliable and autonomous on the order of years in an extremely challenging environment and off-season/unplanned maintenance efforts are seldom necessary. The PermaSense wireless sensor networks consists of Shockfish TinyNodes sensor nodes running the Dozer protocol stack (Burri et al., 2007) implemented in TinyOS (Levis et al., 2005). The sensor nodes are integrated on a custom Sensor Interface Board (Beutel et al., 2009) with power management, data acquisition, storage and interface protection functionality. The analog data acquisition frontend is built using a 16-bit resolution and 8-channel Σ-∆ analog to digital converter (Analog Devices 5 AD7708) and an external precision voltage reference. The ADC is controlled by software running on the MSP430 microcontroller of the TinyNode. The data acquisition operation for both single-ended and differential measurements is configured with a static, periodic sampling rate strictly interleaving with networking operations, in our case 120 s. Other digital sensors, e.g. on-board system health, weather station, digital pressure and temperature sensors can be attached as well using a digital bus interface. The data from the sensor nodes is transferred using the Dozer ultra low-power multihop networking protocol 10 stack (Burri et al., 2007). Data is forwarded to a central data sink, a base station, connected to the Internet with a period of 30 s. In cases of network congestion or loss of connectivity, e.g. due to excessive snow build up or base station failures, data is kept back on local storage on every node using a mechanism called backpressure. For this a 1 GB non-volatile Flash memory storage (SD-card) is integrated on every node. With a power envelope of about 150 µA these wireless sensors have been in continuous operation in the field for periods up to seven years based on a single D-size LiSOCl 2 cell (SAFT LSH-20, 13 A h), 15 although due to maintenance and upgrading activities, in practice the typical operational time on location for a single node is shorter.
Similar to the backpressure mechanism on every sensor node, the base station also contains a local database for intermittent data storage in case connectivity to the database is lost. For reasons of power efficiency the sensor network does not support synchronization to absolute reference time (e.g. UTC) but relies on local 1-second granularity time keeping. The local times-20 tamp of every data sample generated on a sensor node is propagated through the Dozer network and based on the arrival time of each packet at the base station (a Linux system supporting time synchronization to a global reference) the generation time of the respective data sample is calculated using the method of "elapsed time of arrival" (Keller et al., 2012a). Since the forwarding network uses a dynamically changing topology it can happen that data is received out of order with respect to timing at the base station. Because of inevitable drifting behavior of all local clock sources and due to intermittent losses of end-to-end 25 connectivity between nodes of the sensor network as well as on the TCP/IP networking segment slight jumps in the timing can occur (a detailed analysis of the network performance is available in Keller et al., 2011Keller et al., , 2012b. Nevertheless, these effects are not of concern with respect to the long-term nature of the processes observed (diurnal to seasonal behavior). For the user of this data it only matters that on accessing the online data streams on the online data portal in real-time, different timing information exists for every data sample referring to the estimated generation time, the time of arrival at the base station and 30 the time of storage in the data base respectively and that very recent data may still be incomplete (out-of-order arrival with respect to time). Once data has been downloaded, quality checked and possibly also downsampled using the tools discussed in Section 3.4 and supplied alongside with the data in this paper, possible timing artifacts are no longer of concern.
3.2 Low data rate sensor integration The basic sensor used in combination with these wireless sensor nodes are temperature sensors (NTC thermistors) and fracture dilatation sensors (crackmeters) in different configurations ranging from single channel configurations to multiple channel configurations, e.g. 2x crackmeters and 1x thermistor (see Figure 2b) attached to a single wireless sensor node using 3x single ended ADC channels, a half-bridge resistive divider with precision reference resistor and conversion after the Steinhart- 5 Hart equation. A special configuration used are the rigid PermaSense sensor rod and thermistor chain (see Figure 3). These macro-sensor assemblies incorporate multiple thermistors as well as reference resistors, an internal multiplexer circuit allowing to sense at multiple locations (depths) simultaneously housed either in a rigid glass fiber reinforced tube (sensor rod) or located inside heat-shrink tubing and cable segments configured to length as desired. Two variants exist: (i) the original 12 mm 4channel sensor rod that additionally incorporates four electrode pairs allowing to measure resistivity at different depths and 10 (ii) the revised 20 mm sensor rod that is designed without resistivity electrodes but rather in a configurable setup and using metal rings for better thermal coupling to the rock. Both configurations require a 1 m deep hole to be drilled. This most recent design is configurable with respect to the amount of sensors and the sensor depths allowing to manufacture assemblies that are compatible to commercially available units such as the UMS TH3 sensor rods that needed to be replaced as this unit is limited in its measurement range below −20 • C and furthermore requires a lot of power to operate making it unsuitable for long-term 15 monitoring.
Wireless L1-GPS sensor nodes equipped with an additional 2-axis inclinometer for the detection of terrain movement (Wirz et al., 2013) have been developed using the same principle as outlined above (Buchli et al., 2012). Only here GPS data, specifically the RAW output of the satellite observations constitutes the actual sensor data. Environmental forcing, e.g. ambi-

Several Thermistors and Electrodes
(at different depths)

High data rate sensor integration
A number of sensors that are not suitable for integration in a low-power and low-data rate sensor network and that typically come ready to deploy with a standard communication interface (e.g. USB, Ethernet) have been integrated into the field site 5 as well. In order to minimize cabling these sensing systems (e.g. a DSLR camera, high-precision GNSS reference receiver, seismic data acquisition) have been integrated with a Wireless LAN router and facilities to monitor and control power (switch on/off both the sensing system and WLAN from remote). Using a mix of local and remote directional link-based WLAN connectivity between the Internet and the instruments on the field site is established based on a WLAN access point located at the cable car station of the Klein Matterhorn 3883 m a.s.l. about 6.5 km away where the network is attached to a local Internet 10 service provider using fiber.

Data management infrastructure
Care has been taken that all data collected are structured and stored in a coordinated fashion allowing reproducible research results and re-use of data in different contexts and in future projects. Also flexibility with respect to extensions (new sensor types), support of different data rates, metadata integration and life-cycle management were taken into account. The data 15 backend is implemented using a data streaming middleware where a dedicated processing structure called a virtual sensor is responsible for processing a specific data type, e.g. one virtual sensor for temperature measurements and another virtual sensor for images. Complete processing chains, can be implemented by concatenating virtual sensors either within the same instance of the Global Sensor Network ( GSN Aberer et al., 2006) or also across multiple instances of GSN. In our case, data is processed and stored in two concatenated instances of GSN: a private instance only accessible internally for primary, unprocessed data 20 (green database instance in Figure 4) and a public instance for secondary, processed data and publishing this data via web  Matterhorn. The private GSN server receives the data, which are stored in a primary database. Data are passed on to a public GSN server where they are mapped to metadata (positions, sensor types, calibration, etc.) and converted to convenient data formats. Finally, data are available for download and analysis using external tools.
frontend to the user domain, i.e. the Internet (blue database instance in Figure 4). A visualization tool provides up-to-date key graphics (Keller et al., 2012a) on a web frontend where all all data can be accessed online at http://data.permasense.ch. Online data can be accessed using an Internet browser (see Figure B1) or using web queries (see Appendix B).
In this system all data of one specific data type and processing stage is kept in a single data structure with the virtual sensor acting as its interface, i.e. all data of a specific type is kept in this respective data structure irrespective of time and 5 location. The processing chains contain steps for the mapping of device IDs, sensor type and sensor IDs to positions for the respective time periods, applying the correct unit conversion functions according to the sensor type defined, decomposition of more complex data types (multiplexed data) into user-friendly data types and aggregation of data. Each instance of a virtual sensor is mapped to a unique data structure, e.g. a dedicated table on a MySQL database server. Data types with very large amounts of (binary) data, e.g. images are stored directly on a networked file system and only a reference to the respective file 10 is stored in the database. With this two-step data management pipeline consisting of a raw data ingress, dump and store in the first instance as well as multiple processing steps as outlined in the second instance it is assured that all data transactions are consistent, transparent, traceable and verifiable. Should corrections to the data be necessary, e.g. by inclusion of further metadata, correction of metadata or the integration of alternate processing methods they can be applied by simply re-running the respective data from the private primary repository to the second instance with the modifications in place.
In order to consistently manage data of the field site a set of rules has been defined: -An individual protocol sheet is used for each intervention (field work day) where all noteworthy items are recorded (installation, maintenance, removal) 5 -Sensor interventions on site take place at different times for each position. To simplify things, the whole day of an intervention is typically assumed as "invalid data".
-All sensor devices are mapped to a distinct position ID. The mapping contains to-from information, the device id (possibly MAC address), sensor type and calibration data.
-All data from a specific data source (sensor type) is kept in an individual data structure. Queries are typically made per 10 data type and position ID.
-Detailed circumstances (crackmeter angles, thermistor depth) are recorded using auxiliary data formats: text files, excel files or photographs.
As described earlier the data ingress from the base station on the field site is based on a local database on the base station that allows to delay data transmission in cases of loss of connectivity or server outages. In the first years of the deployment 15 this functionality did not yet exist and therefore a (then significant) data gap from June to August 2009 is visible in some of the thermal and crackmeter data due to a failure in the cooling system of the server room and a longer outage of the server system. With hindsight it must be said that this outage event, that had nothing to do with the actual field site instrumentation, exemplified in an extraordinary way the need for tight integration and synchronization of storage resources at all levels of a networked sensing system. 20 4 Detailed field site setup and description of the primary data products This section gives an overview as well as details of the main sensor setup installed at Matterhorn Hörnligrat and describes the data provided with this paper. Table 1 provides an overview listing of the main sensors used grouped by sensor type including their approximate period of operation, units derived, measurement interval and key sensor characteristics. Table 2 and Figure 5 give a detailed listing of the location specific instrumentation detailing the number of sensing channels and sensor 25 types available at each position. For every sensor type used, a detailed discussion of the specifics of each sensor type as well as installation and location specific information is given in the remainder of this section. Finally, Figure 6 gives a graphical overview of the data availability for all data products contained in this paper.
As described in Section 2 and also visible in Figure 6, the sensor setup at this field site has continuously grown over the years.
There are only few data gaps. The data yield and reliability of the measurement systems has surpassed expectations. In a few 30 cases (Position 2 -rockfall, Position 12 -sensor malfunctioning from initial installation) sensing positions have been retired but in general agreement exists that the sensor locations are well planned and selected and that the measurements obtained are representative for each respective location. For the sake of completeness it must be said that a few other sensor placements exist(ed) but due to their experimental nature and/or instability they are not part of this publication.

Weather station data 5
Since 2010 a local weather station based on a Vaisala WXT520 compact all-in-one weather instrument is installed on-site to obtain a more detailed weather data record comprising ambient air temperature (see black line in Figure 7), air pressure, relative humidity, wind (speed and direction) and precipitation. This has been extended with a 4-component net radiometer Kipp & Zonen CNR4 in the summer of 2015 (see green line in Figure 7 for shortwave radiation in). The net radiometer is installed without capabilities for ventilation and heating. The WXT520 is capable of heating the rain and wind sensor but for 10 practical reasons this feature is only enabled when enough power is available which typically corresponds to good weather periods and turned off especially in prolonged bad-weather periods. Both instruments have been vendor calibrated and the  respective calibration data is applied in the data conversion procedures as advised by the manufacturer. It is well known that it is not straightforward to measure present weather conditions in such a hostile and exposed location, high up on the ridge of a 4000 m peak. Therefore this data must be treated with some caution. There are more data outages as with our other sensors.
Clearly an instrument such as the Vaisala WXT520 designed to measure liquid precipitation (with the principle of counting and integrating over the impacts of droplets on the sensor surface) is neither designed nor capable of measuring solid precipitation 5 in any form. Further, the Vaisala WXT520 has been operated in different modes (interval vs continuous sampling) which resulted in different maximum/minimum wind velocity data. Also the application of a net radiometer on a high-alpine rock ridge is far from any WMO compliant sensor setup. Although in parts only indicative, the data obtained from these sensors is very valuable as it is local to the site and exhibits all the small scale local and temporal variability that regional models extrapolating from national service weather data cannot capture, e.g. regular local cloud build up on the mountain slopes in the 10 summer's late afternoons, detailed onset timing of local weather changes etc.

Ground temperature
Ground temperature data are recorded at different depths (ranging from near-surface, which refers to a depth  near-surface temperature, two different major types are used to measure temperature at different depth: on the one hand sensor rods are drilled in the rock and on the other hand thermistor chains are deployed in fractures (see Table 1). All thermistor systems used have been calibrated using a single-point calibration scheme at 0 • C. The main characteristics of the four different temperature measurement devices used are given in the following: 1. PermaSense sensor rod 12 mm: YSI-44006 NTC thermistors, interchangeable tolerance ±0.2 • C, Drift @ 0 • C over 100 months <0.01 • C Table 2 shows which temperature sensors are installed at which position, whereas Table 3 the depths of the thermistors.
15 Figure 8 shows exemplary hourly rock temperatures measured at 10 cm and 85 cm depth and mean annual rock temperature at 85 cm (MAGT_85cm) for years with more than 98% data availability.

Ground resistivity
Electrical resistivity tomography (ERT) is a common geophysical method to characterize the shallow subsurface (Daily et al., 2012). ERT has successfully been used to observe temporal and spatial variations of moisture movement during freeze-thaw cycles in solid rock faces (Sass, 2004(Sass, , 2005 and in solid permafrost rock walls in short- (Krautblatter and Hauck, 2007) and long-term (Keuschnig et al., 2017) measurement campaigns.

5
The PermaSense sensor rod 12 mm is designed with four electrode pairs with a distance of a centimeter each that couple with the rock electrically using conductive foam pads (see Figure 3). In contrast to ERT-surveys, here the contact resistance is directly added to the rock resistance (serial connection). The direct current (DC) flowing through of the rock is measured when excited with a reference voltage (i) at these electrode pairs (at the same depth) in order to gain an indication into the liquid water content and (ii) between electrodes at different depth using and sensor-internal multiplexing unit. The latter configuration has 10 to be interpreted carefully due to the extremely high resistances of this configuration (resistance measurements depend on the contact resistance of the electrodes and on local heterogeneity of the rock between these electrodes). While Table 3 provides the depths of the electrodes for each position, Figure 9 indicates a strong seasonal pattern, which is most likely related to the freezing of the rock. Comparable to the results of a study by Krautblatter (2009), temperature-resistivity gradients for intact porous rock in frozen state here lie in a similar range of about 20 − 40%/ • C cooling (Hasler, 2011).

Fracture displacements
Fracture displacements are measured using Stump/Terradata ForaPot crackmeters. These instruments are very accurate and robust linear potentiometers that are digitized using the wireless sensor nodes described earlier using a resistive half-bridge connection and a single-ended ADC channel per sensor element similar to the temperature measurements. The sensors exhibit a high linearity of ±0.075 % (50-150mm measurement range) and ±0.05 % (200-300mm measurement range) with a resolution 5 better than 0.01 mm and a temperature dependant drift of max. 5 ppm/ • C i.e. 0.25 µm/ • C for a change of 10 • C on a 50 mm range instrument. The devices are specified for operation in −30 to 100 • C. The setup has been validated on site with respect to device interchangeability and long-term stability, the details of which can be found in (Hasler et al., 2012) and the appendix of A. Hasler's PhD thesis (Hasler, 2011). The primary usage of these instruments is to determine displacements perpendicular to a fracture, i.e. the opening and 10 closing movement (see Figure 2a). At select locations multiple crackmeters have been installed in order assess movement both perpendicular as well as parallel to the fracture (shearing) (see Figure 2b and c). In one location (position MH09) a triple crackmeter placement has been installed in order to capture three degrees of freedom of a large buttress detaching from the ridge into the East face. The buttress itself is additionally instrumented with a L1-GPS unit and integrated inclinometer (position MH35) mounted on top of the instable structure. Table 4 lists the details of all crackmeter installations: length of

High-resolution visible light imaging
A time-lapse camera based on a Nikon D300 camera with a 24mm f/2.8D fixed focal length lens has been implemented using the PermaSense base station hardware and a WLAN data link (Keller et al., 2009b). The schedule and parameters for taking pictures can be remotely managed, making it possible to control the camera based on experimental needs. At times when 15 there are no imaging jobs active, the whole system sleeps minimizing overall power consumption to be woken up on request using our low-power wireless sensor network. In this manner, the camera has been operating since 2009, taking many tens of thousands of images from the field site. We have included a selection of images taken at approximately 2-hr intervals at full resolution of the camera (DX format sensor at 23.6 mm × 15.8 mm, 4288 × 2848 pixels, JPEG format). Further images are available in the form of a hand-selected and labeled data set in (Meyer et al., 2018) or directly from the web frontend at http://data.permasense.ch where different resolutions and image formats are also available (select pictures in Nikon RAW (NEF) and/or in variable image resolution).

GNSS raw observation data
In order to assess large-scale movement of individual buttresses of the ridge a number of GNSS sensors are used. A high-  Table 5). 15

Inclinometer data
The wireless L1-GPS sensor systems installed on positions 33, 34, 35 (stations MH33, MH34, MH35, see Table 5) also contain an integrated 2-axis inclinometer based on a MEMS component (Murata SCA830-D07). It is sampled every 120 s, support a ±30 mg offset accuracy over the operating temperature range. The data is transmitted over the wireless sensor network and can be used to assess the rotational movement across the two horizontal axes of the rock mass as well as the height of the mast the 20 GPS sensor is mounted on. For an example of this method see (Wirz et al., 2013(Wirz et al., , 2014 an example of the inclination change combined with displacement derived from daily GNSS position coordinates is shown in Figure 14 for position MH34.

4.8 Further data and related work
In the following, we list different data types and respective sources of data that we know exists and that is closely related to the data collated and documented in this publication and that are not available through a well established (national) data service e.g. weather service or cartographic service. It is a mixture of data that either we obtained by ourselves but is out of scope of this publication either (i) because it is specific to a campaign or purpose, (ii) not mature enough in the sense of quality control 5 and processing or (iii) owned by a related (research) project effort. Nevertheless we take the opportunity to list the data sources we are currently aware of as of writing of this publication. For access to the respective data please contact the data owners given in the references.

Meteorological data
The closest comprehensive meteorological data record relative to the Matterhorn field site are the MeteoSwiss stations Stafel ZER4 and GOR2. If required, these data have to be retrieved from the respective data owners.

Acoustic and microseismic data
Since 2012 a number of different experiments investigating acoustic emission (Weber et al., 2018c), microseismic signals (We-15 ber et al., 2018b) using different instruments ranging from piezoacoustic sensors (>5 kHz), accelerometers (10 Hz-10 kHz) and seismometers (1-100 Hz) have been conducted. The respective data sets for these publications are publicly available and described in detail here (Weber et al., 2018a;Meyer et al., 2018). While the acoustic emission and mid-frequency accelerometer data is highly site specific and experimental, the lower frequency seismometer data is of a more general interest and applicability. Since the end of 2018 this data is being propagated automatically to the Swiss Seismological Service (SED) at ETH Zurich 20 where it is curated and can be accessed online.
Further seismic data originating in a measurement campaign of ARPA VDA, Italy from 2007 to 2012 near the J.A. Carrel hut on the south-east ridge of the Matterhorn at 3829 m a.s.l. is also available (Coviello et al., 2015;Occhiena et al., 2012).

Aerial imaging campaigns
In the year 2013 the UAV company senseFly in collaboration with Pix4D and Drone Adventures performed a demo flight with 25 their UAV drones covering the whole Matterhorn from summit to base. From this campaign a 300 million points 3D pointcloud as well as orthophotos exists. Complementary imaging and scanning products are available by Swisstopo (www.swisstopo. admin.ch).

Terrestrial laserscanning and radar campaigns
Several campaigns using terrestrial laserscanning (TLS) with instruments located both on the Matterhorn Hörnliridge and near the Hörnlihütte below (in 2014, 2015, 2016, 2018) as well as two real aperture radar interferometry (Caduff et al., 2015) campaigns (2015,2016) have been performed. This data can be obtained from the authors upon request.

Wireless network related technical data
A large amount of data concerning sensor status and health, network performance, solar power generation etc. is available 15 over the whole deployment period. The PermaSense wireless sensor network on the Matterhorn constitutes the longest running sensor network for scientific (research) purposes worldwide and arguably also in an extreme environment. This data can be accessed through our online data portal at http://data.permasense.ch but publishing this data within this publication is out of scope.
5 Derived data products, processing and validation methodology 20 For a select amount of the primary data provided with this paper we present derived data products: A number of data sources exhibit very high sampling rates. Depending on the analysis goals these high sampling rates (e.g. 120 s) can be seen as an asset, e.g. to understand small scale, short term process chains but in general when dealing with the whole data set over a decade the gigantic amount of these data constitutes a burden. Therefore, we first introduce a method to downsample these data to reasonable rates in combination with a few data cleaning steps that have emerged as successful out of good practice. Specifi- 25 cally, this method includes (optional) filtering based on sensor-integrated reference resistors (for thermistors and crackmeters), data cleaning based on the manual interventions recorded and the temporal aggregation over 1-hour windows. The resulting data products are file sizes in the order of 100 kB per year rather than 100's of MBs. We provide both a description of the method, the code implementation as well as all input and output data in the context of this paper to allow full transparency and reproducibility. Furthermore by providing a toolset used for all processing steps concerned the reader can adapt processing steps or update the data set independently from future data set updates (living data process).
In the case of the GNSS data the raw GNSS observables are processed to daily positions using double-differential post processing and a local geodetic network as described in Section 5.2. A description of the processing toolset is available in Appendix A2.  Figure 11. Three-step data processing methodology for PermaSense sensor data.
The data stored in the PermaSense GSN public database contains data obtained from sensor nodes after unit conversion.
These data that, we call raw data can be downloaded using a standard web query (see Appendix B). However, since these data are sampled and transmitted independently they do not have a common time stamp and can at times contain discrepancies such as spurious outliers or the response to anthropogenic interventions, e.g. on manual service days. Therefore, a multi-10 step data processing methodology (see Figure 11) is applied, where each step is optional/user selectable (details are given in Appendix A1): Step I: filtering based on reference resistivity data 1 Independent additional electrical resistors are built into the PermaSense sensor chain, PermaSense sensor rod 12 mm and PermaSense sensor rod 20 mm as a means to assess sensor and data integrity (detailed description is given in Section 4.2 and 4.3). After filtering using these reference values, only data with 15 reference resistivity values within a given range (defined in the metadata) are considered for further propagation.
Step II: cleaning using a lookup-table Artifacts in the data either identified manually or systematically known (e.g. on device change interventions) are cleaned using this step. Cleaning operations are delete, set an offset or replace a single or multiple data points.
Step III: aggregation over 1-hour windows For all data types but GNSS data and photographs 1-hour aggregates are calcu-20 lated. For most data types, the aggregation function arithmetic mean was applied. Different aggregation functions were applied to some meteorological data, as an example sum for rain duration or maximum for rain peak intensity. For details, see Table A1 in Appendix A1.

GNSS derived data products
Daily static positions for all GNSS stations are calculated using double-differential GPS post processing based on two different tool chains: using the Bernese GNSS Software (Dach et al., 2015) and the open-source RTKLIB toolchain (Tomoji, 2018). For processing the observables are first collected from the online database and stored in daily observation files with one file per day and position. Double-differencing achieves best accuracy when utilizing the precision final GNSS data products from IGS 5 although other GNSS data products can be used as well. In a final step the position coordinates are converted from WGS84 coordinates to Swiss national coordinates using the online REFRAME conversion service (REST API) by swisstopo. The resulting position data is subsequently uploaded again to the GSN database server from where it can be queried. The geodetic datum of all daily position data is CH1903+/LV95 with the reference frame Bessel (ellipsoidal). After post-processing data for a required amount of days, position data for each position is collated in a single file per position and a number of standardized 10 graphs are generated (see Figure 12). Apart from the raw GNSS observations in the form of daily RINEX 2.11 files we provide the calculated daily positions for both processing toolchains as described above. Further, we provide the scrips and configuration files used to run the opensource RTKLIB toolchain both from prepared RINEX files and from the online data from our database (see Appendix A2).
Double-differential GNSS processing (Teunissen and Montenbruck, 2017) is based on data obtained in a common observation 15 interval from a station pair. Positions for the so-called "rover" can be calculated with high accuracy under the assumption that the "reference" station location is quasi-stationary and that observations from both stations are subject to similar perturbations.
In practical application of this technique care should be taken that the baseline distance between any station pair is short, the field of view to the satellites (horizon) is similar and that a station pair be located in the same altitude regime. Main quality indicators of the input data (GNSS observables) are the number of visible satellites, the signal-to-noise ratio and the observation duration. For the derived data products the ratio of fixed ambiguities as well as the standard deviations per coordinate axis are the key indicators. to showcase some selected data in a visual format. We are only giving a brief introduction and interpretation in the following.

Cross-validation of different sensor data: Examples
Detailed analysis using further methods, especially by leveraging correlation methods that allow to combine data from different sensor types, should be applied to this data, but this is clearly beyond the scope of this paper. 10 In Figure 13 we showcase three types of data in a format suitable for the analysis of frozen ground: Fracture displacement measured using a crackmeter, rock internal resistivity and relative displacement measured using GNSS side-by-side and plotted against temperature, the different years are color coded in order to understand the behavior over time. The data shown originates from four different sensor types at three different locations. All three plots show freeze-thaw related processes that repeat each year as well as an irreversible kinematic component that dominates in summer when temperatures in the rock wall are well 15 above zero.
Similarly, stepwise displacements can be seen when plotting GNSS derived daily positions and a co-located inclinometer on a conventional plot using time on the X-axis (see Figure 14). The first thing to note in this plot is the fact that different sensors and their resulting data types exhibit significantly different error patterns. Here, although the displacement is only on the order of millimeters, the GNSS derived displacements are much more accurate/stable than the inclinometer data that seem 20 to be heavily influenced by present weather conditions, e.g. wind. Over winter periods, the displacement is negligible while the inclinometer raw data apparently relaxes. With the onset of the snow-melt period, an acceleration takes place that can be seen both in the GNSS data as well as the inclinometer. This acceleration continues until late fall. The exact timing of this behavior is known from in-depth analysis of the crackmeter data at Matterhorn (Hasler et al., 2012;Weber et al., 2017).
In the case of the GNSS positions at Matterhorn Hörnligrat all rover positions MH33-MH40, MH43 (the L1-GPS systems) 25 are calculated relative to the two-frequency high-performance GNSS receiver located at MH42/HOGR. However this reference location is also exhibiting significant movement as it is positioned on the top of the buttress between the detachment zone in    Figure C10).
Similarly calculating velocities or aggregate displacements using a simple or more complex method (Wirz et al., 2014) is at the discretion of the data user.

Scientific results based on Matterhorn Hörnligrat data
Data over the period 2008-2011 were the foundation of A. Hasler's PhD thesis (Hasler, 2011) that investigated the thermal and kinematic regime in steep bedrock permafrost for the first time to this extent and level of detail with important contributions 5 to the spatial variability of the thermal regime (Hasler et al., 2011b) and kinematics (Hasler et al., 2012) concluding that enhanced movement in summer originates from hydro-thermally induced strength reduction in fractures containing perennial ice. This hypothesis was later supported when further data became available over a longer monitoring period (Weber et al., 2017). Further, in the wider context of rock slope stability assessment, a new metric was proposed to quantify irreversible displacement of fractures based on the statistical separation of reversible components, caused by thermo-elastic strains, from 10 irreversible components due to other processes (Weber et al., 2017). With the addition of acoustic emission and microseismic sensors to the field site S. Weber's PhD Thesis (Weber, 2018) focused more on structural aspects and the characterization of micro-seismic response to fracture events (Weber et al., 2018c) and on ambient vibrations (Weber et al., 2018b) with the following major findings: (1) A significant amplification of micro-seismic signals in the frequency band 33 − 67 Hz was found.
Filtering in this specific frequency band enables a more reliable detection of fracture events, which is a prerequisite for rock 15 slope stability assessment and early warning.
(2) The characterization of the site specific seismic response based on ambient seismic vibration recordings suggests that the temporal variations in resonance frequencies are linked to the formation and melt of ice-fill in bedrock fractures.

Code and data availability
The data set (doi.pangaea.de/10.1594/PANGAEA.897640, Weber et al., 2019 published with this paper contains data from 5 the start of measurements on July 25, 2008until December 31, 2018. An overview on the structure, file types and size of the data sets, both for the raw primary data and derived data products is given in Table 6. Furthermore the data set also contains the key metadata file for the Matterhorn field site: matterhorn_nodeposition.xslx. Annual updates of this data set are planned (living data process). Using the toolset described in Section 5 and using the online repository at http://data.permasense.ch (see Appendix A1 for details) the data user can also create custom updates of the data set independently.
10 Table 6. Structure, description, formats and sizes of the data set components. The data sets as well as the toolset (code) for preparing, processing, validating and updating the data contained in this publication are available through the following providers and data links:

Conclusion and Outlook
When reflecting on the past ten+ years of development and operation it is fair to say that the promises of distributed wireless systems have delivered unprecedented detail and quality with respect to data. But on the other side the complexity and requirements for mastering increasing degrees of freedom increased as well. What has been especially troublesome at time was the sheer amount of data. Managing and especially the effort for devising a suitable data management system architecture including 5 implementing workable and sustainable solutions has been greatly underestimated. There are no quick answers, make-or-buy decisions are frequently re-visited and there is no ready-made kitting that can be implemented swiftly. Since we believe that this present publication and it's related data set are already large and complex the acoustic emission and microseismic (AE/MS) data from Matterhorn as well as the terrestrial laserscanning and radar interferometry data are not included although it constitutes an integral part of the observations made at this field site. Parts of the AE/MS data has been published separately as we 10 have indicated earlier, but putting all this into a single publication/data set would have simply been overwhelming.
The PermaSense data set from Matterhorn Hörnligrat is the largest, most fine-grained and diverse data set available for permafrost research worldwide. Remarkable about this data set is not only it's duration but also the diversity and density of measurements. The decade+ of interdisciplinary research summarized here shows in an exemplary way how modern (wireless) technological advancements enable new science and the related breakthroughs. The data described here is multi-facetted, 15 exceptionally rich and therefore constitutes a substantial foundation for further research, e.g. in the area of methodology development, the development of process models, comparative studies, assessment of change in the environment, natural hazard warning and preparing for adaptation. Updates to the data set are planned (living data process) but independent of that the user can obtain updates independently using the toolset provided with this data set. Apart from flexibility, this allows also for maximum transparency and reproducibility of the data presented in this paper. 20 Opportunities for future work exist in a multitude of ways and we are only highlighting two directions here: Bringing together the data presented in this paper with data from our colleagues in Italy (ARPA VDA, Matterhorn summit, Carrel ridge, Cima Bianche monitoring site) and SLF/WSL + PERMOS (permafrost boreholes on the Swiss side) allows to obtain further detail over a large span of altitude regime of the European Alps as well as the peculiarities of north-vs. south-facing exposition.
Comparative studies to other similar sites, e.g. Aiguille du Midi, Chamonix, France where the altitude and climatic forcing is Appendix A: Toolsets for generating/processing the derived data products Code for the management and processing of data associated with this manuscript is available at https://doi.org/10.5281/zenodo.2542714, . It contains both a Python3 toolbox for downloading and processing primary data from the online web service at http://data.permasense.ch as well as scripts for post-processing GNSS data using the open-source tool RTKLIB (Tomoji, 2018). Detailed information how to run these tools is given in the README files therefore only a brief synopsis is presented 5 here.

A1 Filtering, cleaning and aggregation toolset
The GSN data management toolbox  is implemented in Python3. It allows to: -Query data from PermaSense GSN server and save it locally as csv-files, -Reload the locally stored csv-files, 10 -Filter according reference values if available, -Clean data manually if needed, -Generate 60-minute aggregates using in principle arithmetic mean (exceptions for weather data are shown in Table A1), -Export yearly csv-files for each position/location, -Generate standard plots for all positions/locations as sanity check and

15
-Query images from PermaSense GSN server and save it locally as jpg-files.

A2 GNSS post-processing toolchain
The open-source RTKLIB toolchain (Tomoji, 2018) is a popular tool for processing GNSS data. It consists of a number of binary tools that can be used both in cmd-line mode and in combination with a GUI as well as the respective configuration files. In order to automate the processing of larger data sets we have developed a small toolchain that allows to prepare all data 20 necessary and calculate double-differencing daily position solutions. In order to use this toolchain an operational installation of RTKLIB is required. For details on RTKLIB please refer to the respective tool documentation. The top-level shell script compute_solution.sh allows to specify a configuration parameter file, several options and the day for which processing is to be performed:  The parameter file specified contains information on the baseline pair being processed, data products used and the exact locations of servers and directories to be used. The latter of which need to be adapted to suit your specific installation. The compute_solution.sh shell script calls further auxiliary programs written in python as well as tools from RTKLIB. The syntax is best explained using an example for computing positions MH42/HOGR and MH33 for the first day of the year 2017: ./compute_solution.sh -p parameter_file_HOGR_ZERM.txt -b -r -c -d -f 2017 01 01 ./compute_solution.sh -p parameter_file_MH33.txt -b -f 2017 01 01 An example of how this toolchain can be used to compute daily positions for all Matterhorn GNSS positions for a given day 5 is shown in the shell script gps_batch_compute.sh that can also be used to automate this process on a compute server.

33
Appendix B: PermaSense online data access For use cases where updates to the data being provided with this paper or direct access to the online database is required we include a short introduction to the web interface and it's query syntax here. The data in the GSN database available at http://data.permasense.ch is organized in data structures called virtual sensors (VS) per deployment (see Figure B1). If there are multiple sensors yielding the same data types, this data is multiplexed into the same VS. Each VS has a unique name: 5 <deployment>\_<sensor type>. For convenient data download the web frontend supports complex queries using the multidata query interface 2 of GSN with the following options: 15 specifies the name of the virtual sensor, time_format specifies the time format of the returned data, field[0] specifies the list of data fields to return and the from, to clause limits the time window of the query. The result of this query is a CSV-formatted file with the requested data, in this case all sensor positions will be reported that produced data in the given time interval. Typically a query for data pertaining to a single position only will employ further limits, e.g. on the field position as follows for a limit to position 3:  The virtual sensor to which the condition is to be applied. All or the name of the vs All c_field[n] The parameter to which the condition is to be applied. All or the name of the parameter All c_min[n] The minimum value of the condition to be met.
-inf or the minimum value All c_max[n] The maximum value of the condition to be met.
-inf or the maximum value All Figure B1. The online data management web frontend at http://data.permasense.ch allows to access all data in real-time. Data are accessed by data type in entities called virtual sensors (right). Selected standard views, e.g. key graphs can be accessed via the tabs at the top.

36
Appendix C: Pictures of the field site and selected instrument details Figure C1. From the south the large detachment scar (light grey rock) to the left of the deeply incised second couloir on Matterhorn Hörnligrat is well visible. Figure C2. North of the detachment zone (light grey colored rock) a small ice field is visible delimiting the strongly fractured topography close to the ridge from the north face.         issues, long-term monitoring strategy, data curation, policy and strategic decisions as well as to Wilfried Häberli for persistent guidance, judicious stewardship and friendship through all the years. Reviews from Paolo Pogliotti, and an anonymous referee provided valuable comments that helped to improve the paper substantially. Finally, we thank the handling Editor Kirsten Elger for constructive feedback and suggestions.