Publications
Overview of papers, conference contributions, and research outputs. Fetched from NVA.
Automated UAV systems for geohazard monitoring: case studies from the Supphellebreen icefall (Norway), the Skjøld instability (Norway), and the Blatten landslide (Switzerland)
Natural Hazards and Earth System Sciences Discussions (NHESSD) (2026)
Abstract. This study presents the first systematic field evaluation of dock-based UAV (Uncrewed Aerial Vehicle) systems for geohazard monitoring in mountainous terrain. We tested their potential across three different environments: (1) a fast-moving glacier icefall (Supphellebreen, Norway), (2) an unstable rock slope (Skjøld, Norway), and (3) a post-failure landscape resulting from a catastrophic rock-ice avalanche (Blatten, Switzerland). Effective hazard management requires timely detection of displacement patterns and terrain change. To address these issues, we introduce an automated workflow integrating multitemporal UAV dock data acquisition with an end-to-end processing pipeline for displacement field generation and change detection. The results show that this workflow has the potential to provide data at centimetre-level accuracy before, during, and after hazard events, supporting both precautionary risk assessments and timely decision-making in critical phases of potential hazard evolution. Wider adoption will depend on supportive regulatory frameworks, reliable power and communication infrastructure, and sufficient expertise to ensure effective operation, maintenance, data interpretation and risk management. Overall, dock-based UAV systems represent a significant technological advancement in efficient geohazard monitoring, facilitating rapid response in critical situations, thereby contributing to increased resilience of communities living in vulnerable mountain environments.
Natural fracture ventilation influencing the behavior of rock slope instabilities
Nordic Geological Winter Meeting
Networks of open bedrock fractures are a common feature in unstable rock slopes. Such fracture systems influence the subsurface thermal regime by allowing air circulation and water infiltration. Ground temperature conditions are in turn a crucial factor controlling slope stability. Mapping subsurface fracture systems in unstable rock slopes therefore remains one of the biggest challenges in assessing the risk such slopes pose to nearby inhabitants and infrastructure. The Stampa rock slope (Aurland, Norway) is a complex instability with an estimated total volume exceeding 800 Mm³. Although a collapse of the entire slope is considered unlikely, unstable rock sections along an up to 100 m high rock cliff represent potential sources for partial slope failures. This cliff marks the transition between the scree slope below and the gently sloping mountain plateau above, which hosts morphological depressions, graben structures, and open tension fractures where air flow at vents has been observed. To investigate the underground fracture system, we have systematically mapped vent locations in the field aided by UAV-based optical and thermal imaging. We equipped 11 fracture vents with air temperature loggers, three of them with rock surface temperature loggers, and two of them with sensors to monitor velocity and direction of air flow, as well as radon content. These continuous monitoring data are complemented with occasional in-situ air flow and temperature measurements at fracture vents, in addition to radon (222Rn) and thoron (220Rn) surveys using alpha track detectors. Our results show that there exist several complex ventilation systems connecting vents at different elevations both on the mountain plateau and in the rock cliff. Chimney ventilation is probably most important and is predominantly controlled by the outside air temperature, since the direction of air flow depends on the season. In winter, when outside temperatures are low, air is sucked in at lower elevated vents and flows out at higher elevated vents; in summer, the pattern reverses. The flow rate is controlled by the temperature gradient between outside air and air in the bedrock fracture system, and reaches up to 1 m3/s. 222Rn concentrations are strongly dependent on the air-flow rate and are generally highest (up to 50 000 Bq/m3) when the flow rate is low, but they are also related to the rock-atmosphere contact area which depends on ground water level and the extent of ice in subsurface fractures. Stable subsurface air temperatures at or below 0°C indicate the existence of sporadic extra-zonal permafrost mainly in the rock cliff below the mountain plateau. There, the bedrock reaches its highest temperatures in late autumn/early winter which coincides with enhanced deformation of unstable rock sections along the cliff. Climatic warming leads to changes in bedrock ventilation systems, warming ground and thawing permafrost, and thus influences slope stability.
A Fine Line: Explaining Landslide Forecasting with Dynamic Rainfall Thresholds
Western Norway University of Applied Sciences PhD Symposium 2025
Rainfall induced landslides globally constitute a major hazard in mountainous areas, threatening infrastructure and human lives and are expected to occur more frequently as a consequence of climate change. Recent studies utilising Machine Learning instead of physically based models or empirical thresholds have been focusing on established deep learning architectures, leaving modern architectures under-explored. The apparent reluctance for adopting modern deep learning methodologies could be linked to the inherent difficulties in interpreting these models’ decisions, a crucial component for operational landslide early warning systems. We propose the use of Kolmogorov-Arnold Networks (KANs) for landslide prediction based on precipitation time series from a globally available satellite product. In addition to the inherently interpretable activation functions of KANs, we propose the usage of Dynamic Rainfall Thresholds (DRT) as a visual interpretation tool for the model.
Autonomous UAVs for Monitoring Glacier Dynamics and Hazards: A Case Study from Jostedalsbreen, Norway
EGU General Assembly 2025
Monitoring glaciers is essential for understanding their response to climate change, managing freshwater resources, and mitigating geohazards such as icefalls and glacial lake outburst floods (GLOFs). Traditional glacier monitoring techniques often face challenges connected to limited spatial and temporal resolution and logistical constraints in hazardous terrain. These challenges are especially pronounced for steep and fast-moving glaciers with large surface changes and high velocities. In such settings high temporal and spatial resolution data are essential for capturing rapid surface changes and understanding glacier dynamics.We introduce the potential of autonomous unmanned aerial vehicles (UAVs) operating from stationary drone docks as a novel, flexible, and cost-effective solution for glacier monitoring. We tested a DJI Dock 2 at Flatbreen and Bøyabreen, two outlet glaciers of the Jostedalsbreen ice cap in Western Norway. We captured high-resolution aerial imagery for photogrammetric mapping, conducted at customizable intervals (hourly, daily, weekly). These datasets enabled the generation of multitemporal point clouds, digital terrain models and orthophotos. To derive surface velocities and detect changes over time we used the 3D point cloud analysis algorithm M3C2 and 2D feature-tracking methods.Preliminary findings revealed that autonomous UAVs can monitor surface changes and velocity patterns effectively with a high temporal and spatial resolution. Surface velocities for both glaciers ranged from 0.4 to 1.5 m per day, with higher rates observed in steeper sections of the glacier. The data offers unique insights on short-term processes, including acceleration phases, crevassing, the collapse of subglacial cavities and several significant icefall events. The results demonstrate a level of detection of 2-4 cm, which allows for the identification of subtle changes at cm-scale. Integrating autonomous UAVs into existing glacier monitoring frameworks represents a significant advancement in data collection by improving spatial and temporal resolution and time efficient workflows through automation in data collection and post processing.This study highlights the feasibility and effectiveness of autonomous UAVs for near-continuous glacier and geohazard monitoring, particularly valuable in inaccessible or dangerous environments. We demonstrate the potential of autonomous UAVs to track both long-term glacier dynamics and short-term changes. This capability enhances process understanding and provides a robust foundation for developing UAV based early warning systems for glacial hazards. While challenges remain, particularly in difficult weather conditions, low visibility, and regulatory compliance, this innovative approach demonstrates substantial potential for monitoring, supporting effective risk management in regions vulnerable to glacial hazards.
Can natural air and radon fluxes help us mapping subsurface fracture systems and understanding the behaviour of deep-seated unstable rock slopes?
NGF Winter Conference
Mapping subsurface fracture systems and potential sliding planes in deep-seated unstable rock slopes is one of the biggest challenges in assessing the risk such an unstable rock slope poses to inhabitants and infrastructure. We are testing the use of natural air and gas fluxes for studying these inaccessible subsurface fracture systems and for gaining information about slip behavior of the rock mass. The radioactive gas radon (222Rn) naturally exhales from soil and bedrock and, due to its more than three days half-life, may travel far through open fracture networks before reaching the surface. Travel distance and speed depends on the ventilation regime and, if ventilation is poor, radon may build up in underground cavities. With ground deformation or movement, radon may escape from underground cavities leading to an increase of radon exhalation from fractures. In combination with thoron (220Rn), which has a much shorter half-life, the gas fluxes may shed light on the age of air exhaling from fractures. We study gas fluxes in the context of natural air ventilation through fracture systems. The characteristics of these air fluxes are strongly connected to the interplay between outside air temperatures and the ground thermal regime, which in turn can have a significant effect on slope stability. The complex unstable rock slope “Stampa” in Aurland, Western Norway, has an estimated total volume of over 800 Mm3. The upper part of the instability can be characterized as a sloping mountain plateau, featuring morphological depressions, graben structures and open tension fractures. The latter is a clear sign for recent activity, although displacement seems minimal, and a sudden failure of the entire slope is unlikely. We have instrumented selected fracture vents with air flow and radon sensors and are measuring subsurface air temperature at 11 vents. The continuous monitoring setup is complemented with sporadic alpha track radon surveys, measuring radon exhalation at fracture vents, and UAV surveys using both optical and thermal imaging. First results show that air flow directions follow a seasonal pattern and that outflowing air at vents has stable cool temperatures. Considering the so-called chimney effect, we can make the following assumptions: (1) the lack of a systematic distribution of summer and winter vents (air flow out) indicates a complex subsurface fracture system with several smaller sub-systems, (2) stable 0°C subsurface air temperatures at some summer vents may indicate sporadic extra-zonal permafrost influencing the slope’s stability, and (3) in order to be able to use radon and thoron concentrations to make assumptions about the slip behavior more data needs to be collected. The gas concentrations are strongly dependent on air-flow rate and rock-atmosphere contact area which, in turn, depends on ground water level and the extent of ice in subsurface fractures.
Creep in scree deposits at the unstable rock slope of Skjøld, Vang Municipality, Norway
Geofaredagen
This study investigates the complex and unstable rock slope Skjøld. The aim is to improve our understanding of the geomorphological, structural and meteorological factors that control the creeping displacement of its unstable scree deposits, which are highly variable in both space and time. Warming and subsequent degradation of permafrost might represent an important process that influences slope stability in the area. The complex displacement patterns in the scree deposits may also mask critical, deeper deformations in the unstable slopes, in particular when relying on measurements of surface displacement. Ground-based radar surveys conducted by the Norwegian Water Resource and Energy Directorate (NVE) reveal line-of-sight displacement rates of up to 4.4 mm/day in June 2020 and 3 mm in four days in August 2024. These differences in displacement rates reveal seasonal changes, which may be controlled by both precipitation, infiltration, and ground temperature. Preliminary results derived from airborne laser scans and multitemporal UAV surveys show horizontal displacement rates of 0.2-0.4 m/yr in the most active areas. To investigate the potential occurrence of extra-zonal permafrost, temperature loggers were installed at the site. Preliminary results from June to August show a relatively close correlation of ground temperature to air temperature, both of which remained above 0°C, indicating unfavourable conditions for permafrost. In September 2025, an ERT profile was collected on the site in order to reveal permafrost areas, estimate the thickness of the scree deposits and to delineate structural weaknesses and potential slip planes in the subsurface. Preliminary results reveal zones with very low resistivity values, which could indicate bedrock, ground water or frozen ground. Our preliminary results do not indicate a permafrost driven instability, but rather structurally or hydrologically controlled displacement. Further measurements and data analysis will give a more complete picture.
eXplainable AI‐based causal discovery for analysing unstable rock slopes
Norsk Geologisk Forening (NGF) Vinterkonferansen 2025
Analysing unstable rock slopes and understanding their drivers is a crucial element for accurate risk assessment and designing appropriate mitigation measures. Traditional approaches, based on statistical analyses and domain knowledge, result in educated guesses about the underlying mechanisms that drive slope displacements, without drawing a definitive conclusion. eXplainable Artificial Intelligence (XAI) based Causal Discovery aims at inferring causal relations between various factors that influence movements on a rock slope from purely observational data. Causal relations between factors describe true causal influences in the underlying process, rather than statistically correlated co- occurrences. The aim of causal discovery is to construct a causal graph that links the different system variables by ways of their causal interactions. We present a novel approach to conduct XAI based Causal Discovery on rock slope displacement data, the Multi-Window Causal Discovery (MWCD), adapted to the unique challenges of geotechnical observation data. Our method can be used to reveal information on general displacement drivers and their evolution over time. Using monitoring data from three different sites, including pre-failure displacement and environmental time series, we demonstrate the effectiveness of the approach to enrich statistical and geological analyses. Applied to case study data from the Stampa 4a rock section in Aurland, Norway, MWCD detects a shift from an indirect to a direct effect of both precipitation and infiltration on the displacement over the monitoring period. This corresponds to the final destabilisation before the failure of the rock section. Using monitoring data from the Preonzo instability in Switzerland, MWCD indicates significant changes in the causal connections of the system, from a temperature and precipitation dominated to a mainly gravitationally driven system. Lastly, at the Veslemannen site in Romsdalen, Norway, MWCD is used to construct a causal graph linking the movements on the upper sections of the slope to movements on the lower sections of the slope, emphasising the applicability for spatio-temporal analysis. We show that the use of XAI-based methods for analysing unstable rock slopes can be a powerful addition to traditional geotechnical analysis and risk assessment. An improved understanding of the displacement dynamics can further aid in designing monitoring setups and building more accurate forecasting methods.
Explainable Artificial Intelligence Based Displacement Analysis and Forecasting for Unstable Rock Slopes
General Assembly 2025 of the European Geosciences Union (EGU)
Geohazards such as landslides, rock avalanches or rock falls from unstable slopes can seriously threaten human life and infrastructure. Monitoring unstable slopes coupled with real-time data analyses to assess the risk they pose and mitigate this risk is thus indispensable. Machine learning-based methods for analysing monitoring data recently significantly improved the forecasting possibilities for failure events. However, one major limitation of Machine Learning-based methods is that they primarily provide "Black Box"-models. These models can, for example, transform arbitrary input into a sequence of predictions, albeit without a transparent explanation of how the output is derived from the input. Even though State-of-the-Art Machine Learning often outperforms traditional failure forecasting methods, such as the Inverse Velocity method, this limitation greatly hampers the application of these methods in practice. Recent advances in eXplainable Artificial Intelligence (XAI) have led to the development of the field of Causal Artificial Intelligence. As opposed to many Machine Learning approaches which are based on Deep Neural Networks, XAI aims to offer transparent models that provide explanations for model outputs. We therefore propose a novel forecasting approach based on XAI, leveraging Graph Neural Networks and Kolmogorov-Arnold Networks. Our approach aims to learn a causal model of an unstable slope or one particular section of it, including slope-internal and meteorological factors that can be represented as a graph, visualising cause-and-effect relationships between the variables. As such, our goal is twofold, and we aim at (1) providing insight into the mechanisms driving slope displacement, and (2) using this information for explainable short-term forecasting by selecting only causally related features from all available data. We apply our method to two case study sites for displacement driver analysis and short-term displacement prediction and compare the model performance to recent State-of-the-Art models. Our method not only aligns with but even outperforms existing models in terms of prediction accuracy and offers, in addition, superior interpretability. The proposed framework provides crucial support for geohazard assessment and monitoring network design. Furthermore, the displacement prediction has great potential as standalone predictive network as well as for hybrid failure prediction methods, for example in combination with traditional long-term failure predictions such as the Inverse Velocity method. While developed with medium-scale rock sections in mind, the method may be adapted to larger rock volumes as well as slow-moving mass movements with failure potential in general. The usage of accurate and interpretable prediction models represents a significant advancement, overcoming the transparency issues of models generated by complex Artificial Neural Networks, ultimately contributing to improving Early Warning Systems.
Explainable Artificial Intelligence for Rainfall Induced Landslide Forecasting using Kolmogorov-Arnold Networks and Dynamic Rainfall Thresholds
Third Workshop on the Future of Machine Learning in Geotechnics (3FOMLIG)
Rainfall induced landslides globally constitute a major hazard in mountainous areas, threatening infrastructure and human lives. Current climate projections suggest an increase in extreme weather events as well as landslide occurrences (Auflič et al., 2023, Gariano & Guzzetti, 2016). Accurately predicting rainfall induced landslides is thus of critical importance to employ effective hazard mitigation as well as evacuation measures. Traditionally, physically based models as well as empirical rainfall thresholds based on intensity and duration of rainfall events are used to predict landslides (Berti et al., 2012, Brunetti et al., 2010). The advent of Machine Learning allowed for the integration of static predisposing factors, such as slope angles and soil types, further enhancing predictive capabilities as well as allowing for a larger spatial application (Dal Seno et al., 2024). However, recent studies utilizing Machine Learning have been focusing on established deep learning architectures, leaving modern approaches underexplored (Mondini et al., 2023, Nocentini et al., 2024). The apparent reluctance for adapting modern deep learning methodologies could be linked to the inherent difficulties in interpreting these models’ decisions, a crucial component for operational landslide early warning systems (Gu & Dao, 2024, Y. Zhang & Yan, 2023). Recent work with focus on traditional Machine Learning often includes an analysis of the network behavior based on post-hoc explanations, using for example Shapley values and feature importances (Nocentini et al., 2023, Wen et al., 2025). However, inherently explainable deep learning networks have to our best knowledge not been explored for application in rainfall induced landslide prediction. To this end, we propose the usage of Kolmogorov-Arnold Networks (KANs) for landslide prediction (Liu et al., 2024). KANs have been shown to be competitive in performance to other deep learning-based approaches for a variety of tasks, such as regression, classification and vision (Han et al., 2024, Xu et al., 2024, Bresson et al., 2024, Cheon, 2024). We propose the use of KANs to predict rainfall induced landslides based on precipitation time series derived from a globally available satellite product. The presented model is competitive in performance when compared to various well-established models: Random Forest (Breiman, 2001), Multi-Layer Perceptrons (MLPs) and XGBoost (Chen & Guestrin, 2016). In addition to the interpretable activation functions, we propose the usage of Dynamic Rainfall Thresholds (DRT) as a visual interpretation tool for the model. The inherent interpretability of KANs paired with the innovative DRT interpretation makes the model a suitable choice for critical applications such as landslide early warning systems, granting unprecedented insight into the system abstractions that Machine Learning models learn during training.
Forstå Ustabile Fjellsider med Kunstig og Naturlig Intelligens
geoforskning.no
Multistage 54,000 m3 rockfall (Stampa, Western Norway): Insights from comprehensive monitoring and failure analysis
Landslides. Journal of the International Consortium on Landslides (2025)
Abstract In 2023, a 54,000 m 3 large rock section failed catastrophically from the unstable rock slope Stampa (Western Norway). The failure occurred in a multistage process with two major failure events on July 1 st and 3 rd . In this study, we present a detailed analysis of the pre-failure displacement patterns, the failure mechanisms, and failure events. After exponential acceleration, the base (12,200 m 3 ) of the rock section failed, leading to a further destabilization of the remaining rock column (41,800 m 3 ), which failed 2 days later. The monitoring data includes in situ displacement sensors, a robotic total station, and close range and remote sensing data spanning over 14 years. The rock section showed seasonal displacement patterns clearly influenced by meteorological factors: (1) late spring (May–July) accelerations, controlled by positive temperatures, thawing ground, and meltwater infiltration, followed by (2) summer stabilizations characterized by low displacement, and (3) autumn (October–November) accelerations driven by precipitation events over a longer period, followed by (4) winter stabilizations approaching zero displacement. The deformation rates increased from 0.06 m a −1 in 1991–2019 more than a tenfold to 0.78 m a −1 in 2022 and indicate progressive damage and weakening of the rock section, which increased its sensitivity to rainfall and infiltration. Our findings highlight the importance of long-term, high-resolution monitoring using different, independent sensors, alongside detailed failure analyses, in understanding the evolution of unstable rock slopes. This study contributes to the understanding of progressive medium-scale rock slope failures, aiding in the prediction and mitigation of potential failures.
Decoding Unstable Rock Slopes - Causal Discovery for Natural Hazard Mitigation
Western Norway University of Applied Sciences PhD Symposium 2024
Unstable Rock Slopes pose significant threats to communities and infrastructure in their vicinity. Understanding the factors causing slope deformation and potentially failure, is critical. Traditional approaches to build Causal Models rely on expert knowledge and simulations to construct qualitative cause-and- effect relations between the observed variables in a monitoring setup.Graph Neural Networks can aid in overcoming the limitations of these traditional approaches. They provide a way to understanding the drivers of displacement in a transparent way through eXplainable AI (XAI).
Monitoring displacement patterns, acceleration and failure (July 2023) at the unstable rock slope Stampa, Western Norway
Nordic Geological Winter Meeting
Acceleration in a rock column within the unstable rock slope ‘Stampa’ Aurland municipality, Vestland county
Geological Society of Norway Winter Conference 2023
Displacement patterns, acceleration and failure from the unstable rock slope Stampa, Western Norway
6th World Landslide Forum
July 2023 failure from the unstable rock slope Stampa, Western Norway
Geofaredagen 2023
Multimodal Asynchronous Kalman Filter for monitoring unstable rock slopes
Geomatics, Natural Hazards and Risk 14(1): 2272575 (2023)
Unstable rock slopes pose a hazard to inhabitants and infrastructure in their vicinity, necessitating advanced monitoring methodologies for timely risk assessment and mitigation. Recent geotechnical monitoring techniques often rely on sensor data fusion to enhance forecasting for imminent failures. Our investigation extends beyond a single sensor type to data fusion for heterogeneous sensor networks using a Multimodal Asynchronous Kalman Filter. We illustrate the application of the proposed method on a case study data set consisting of data from an on-site sensor network enriched by remote sensing data. Employing a Multimodal Asynchronous Kalman Filter, we capitalise on the distinct resolutions inherent in each sensor input. The outcome was a combined dataset with a high spatiotemporal resolution. Our approach facilitates the estimation of essential physical attributes for monitored objects, encompassing translation, rotation, velocities and accelerations. The case study site was an unstable rock section of ca. 50.000 m3 in Aurland, Norway, which collapsed as a multi-stage failure in July 2023. Our method can be transposed to various sites with distinct sensor networks, enhancing state estimations for objects on unstable rock slopes. These estimations can significantly improve applications such as risk assessment and robust early-warning systems, enhancing predictions of critical failure points.
Sensor Fusion for Monitoring Unstable Rock Slopes-A Case Study from the Stampa Instability, Norway
EGU General Assembly
The unstable rock slope Stampa is located north-east of the touristic town of Flåm, Norway along the Aurlandfjord and displays signs of post-glacial deformation over a large area and a volume of several million m3. Directly below the rock slope lies the European Road E16, a highly frequented connection between Bergen and Oslo. Two high-risk objects have been identified on the instability, which are currently being monitored continuously by the Norwegian Energy and Water Directorate. The Landslide Research Group at Western Norway University of Applied Sciences uses an object on the unstable rock slope, Block 4a, as a field laboratory for sensor networks. The approximately 5,000 m3 Block sits on a highly fractured base of approximately 40,000 m3 and has recently been moving at speeds in excess of 1 cm per day. Different failure scenarios threaten the European Road under the object and potentially the town of Flåm. Data from an on-site sensor network with a range of instruments such as wire-extensometer, inclinometer, temperature loggers and geophones has been collected over a period of three years and combined with remote sensing data from a robotic total station, ground-based InSAR and satellite-based InSAR with the use of a corner reflector as persistent scatterer as well as weather station data from Stampa. Sensor Fusion has been used to merge the data of the different sensors and exploit the different resolutions of the respective sensors. This led to the development of a data set with high spatiotemporal resolution capturing the physical properties of Block 4a, such as displacement direction and velocity. This approach makes use of complementary sensor data to fill gaps in time series of other sensors, which can be caused by sensor faults or are due to sensor down-times during maintenance. Both the sensor fusion approach as well as filtering of outliers requires expert knowledge about the system in question, which sensor fusion research groups often do not integrate into their analysis. We propose thus a holistic analysis approach at the intersection between data science and geology. Preliminary analyses of the augmented data for Block 4a confirm high displacement rates at the end of 2022. This follows a general trend of acceleration that has been observed over the last three years. Furthermore, the displacement accelerations seem to follow a seasonality, with acceleration phases in spring and autumn, while summer and winter coincide with less movement. Based on the sensor fusion analysis we can identify that rain fall periods in autumn as well as snowmelt in spring have an impact on the block displacement. However, we conclude that precipitation alone cannot explain acceleration phases. Instead, we propose a model based on the combined influence of rain and snowmelt paired with air and rock surface temperature on the slope movement. In combination with a refined sensor fusion process, we expect our work to be transferable and relevant for the monitoring of other unstable rock slopes.
Subsurface fracture ventilation and gas emission and their relation to rock slope deformation
6th World Landslide Forum