Orals: Fri, 2 May | Room N2
Landslides pose substantial risks to both local populations and critical infrastructure in high-risk areas. Numerous technologies have been developed to monitor landslides, resulting in a growing amount of landslide monitoring data, such as very high resolution remote sensing data and in-situ monitoring data. These data have great potential for developing advanced machine learning models for geohazard assessment. Privacy and security issues are raising concerns, hindering the collection of large datasets required for developing powerful machine learning models. However, existing landslide detection models explicitly or implicitly assume that landslide monitoring and mapping data are directly shared on a centralized server. This assumption leads to a gap between data sharing practices and machine learning modeling in landslide detection. To bridge this gap, we leverage a privacy-preserving machine learning model for the landslide detection task. First, a federated learning method is introduced to protect data privacy throughout the modeling process, enabling the development of landalide detection models without the need to share raw data. Second, we introduce a fair incentive mechanism to evaluate the contributions of participants and encourage more data owners to engage in landslide data sharing. Finally, experimental results demonstrate that the proposed framework effectively protects data privacy while maintaining high prediction accuracy. This approach not only facilitates secure data sharing but also enables institutions to develop more robust machine learning models for geohazard assessment, thereby advancing the field of landslide prevention and mitigation.
How to cite: Tang, X., He, L., Yan, X., Ye, X., Dai, K., Novellino, A., Li, H., Heidarzadeh, M., and Catani, F.: Protecting Data Privacy in Landslide Detection Using Privacy-Preserving Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14403, https://doi.org/10.5194/egusphere-egu25-14403, 2025.
In recent years, deep learning models have been used for automated landslide mapping. However, such models often underperform when encountering out-of-distribution (OOD) data (regions or terrain characteristics that are significantly different from those seen during training). To address this issue, we present an automated application powered by Google Earth Engine that constructs hyperlocal machine learning models tailored to specific areas of interest. By defining a limited spatial extent and providing labels specific to the area, our approach mitigates the risk of encountering OOD data, reducing incorrect predictions. The application supports the export of annotated landslide data in both raster and vector formats, enabling users to validate and refine landslide extent. These new high-quality datasets can be incorporated back into existing deep learning models to improve generalizability. With its speed, accuracy, and user-friendly interface, the proposed app aims to facilitate the development of robust landslide identification models, especially in scenarios where data scarcity or geographic diversity poses significant challenges.
How to cite: sharma, N. and saharia, M.: Mitigating Out-of-Distribution Challenges in Landslide Mapping through a Hyperlocal Machine Learning model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10302, https://doi.org/10.5194/egusphere-egu25-10302, 2025.
The earthquake-induced landslide targets in remote sensing images vary greatly in size and are unevenly distributed with many small targets. Achieving a balance between high accuracy, computational capability, and small sample size remains challenging. This study proposes to enhance earthquake-induced landslide detection by developing a new algorithm for remote sensing images based on the C3-Swin-Transformer and Multiscale Feature Fusion-YOLO (TLSTMF-YOLO). Utilizing a feature extraction layer and Swin-Transformer structure captures dependencies and preserves spatial information. Introducing the Convolutional Block Attention Module (CBAM) enhances feature representation. Incorporating a Bidirectional Feature Pyramid Network (BiFPN) optimizes bidirectional cross-scale feature fusion, improving landslide detection accuracy across scales. The training utilizes an AdamW optimizer and cosine learning rate strategy for accelerated convergence and improved speed. Transfer learning applies to Jiuzhaigou and Luding landslide datasets. Experimental results show that the TLSTMF-YOLO model outperforms YOLOv5 and other detection models in terms of precision, recall, and mAP@0.5. Specifically, on the Jiuzhaigou dataset, it achieves a precision of 95.7%, a recall of 89.9%, and a mAP@0.5 of 90.5%. On the Luding dataset, it achieves a precision of 96.0%, a recall of 90.9%, and a mAP@0.5 of 94.5%. Additionally, the frame processing times for the TLSTMF-YOLO model are 6.61 ms and 12.2 ms on the Jiuzhaigou and Luding datasets, respectively, demonstrating superior efficiency and confirming its effective feature extraction and fusion capabilities.
How to cite: Meng, S., Shi, Z., Peng, M., and Glade, T.: Earthquake-Induced Landslide Detection in Remote Sensing Images Using TLSTMF-YOLO, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1478, https://doi.org/10.5194/egusphere-egu25-1478, 2025.
Landslides represent a significant geological hazard, particularly in mountainous regions where ground deformations can lead to devastating impacts on infrastructure, ecosystems, and communities. The Belluno Province, situated in the Veneto region of northeastern Italy, is characterized by its complex topography and geological features, rendering it particularly susceptible to landslide occurrences. To mitigate the risks associated with these natural phenomena, effective hazards mapping is essential. This study explores the integration of interferometric synthetic aperture radar (InSAR) data with multi-temporal inventories to enhance the accuracy and reliability of landslide hazard assessments in this region. By leveraging advanced remote sensing techniques alongside landslide data, this research aims to provide a comprehensive spatial analysis that identifies areas at risk and contributes to informed decision-making in land management and disaster mitigation. To this end, considering slope units, the landslide data delineated using orthophotos retrieved from WMS and WMTS services provided by the Italian national portal, covering the period from 1989 to 2021, were analyzed. The analysis focused on the Cordevole and Alpago regions, located in the Belluno province of the northeastern Italian Alps. These areas were affected by two extreme meteorological events with a return period of over 100 years: the first, a windstorm named VAIA, occurred from October 27th to 30th, 2018, and caused significant damage to the forest cover. The second event took place from December 4th to 6th, 2020, also impacting the region. The findings of this integration not only hold implications for local stakeholders but also enhance the broader understanding of landslide dynamics in similar geological contexts.
Acknowledgement:
This study was carried out within the PNRR research activities of the consortium iNEST (Interconnected North-Est Innovation Ecosystem) funded by the European Union Next-Generation EU (Piano Nazionale diRipresa e Resilienza (PNRR) – Missione 4 Componente 2, Investimento 1.5 – D.D. 1058 23/06/2022, ECS_00000043). This manuscript reflects only the Authors’ views and opinions, neither the European Union nor the European Commission can be considered responsible for them.
How to cite: Bhookya, R., Puliero, S., Floris, M., and Meena, S. R.: Integration of InSAR data with Multi-Temporal Inventories for Potential Landslide Hazard Mapping in Belluno Province (Veneto Region, NE, Italy)., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4168, https://doi.org/10.5194/egusphere-egu25-4168, 2025.
The catastrophic rainfall events of May 2023 in Emilia-Romagna, Italy, triggered over 80,000 landslides and widespread flooding, presenting unprecedented challenges for emergency response and disaster management. This study evaluates the potential of automated landslide mapping using deep learning models, specifically U-Net and SegFormer, to address these challenges in scenarios with limited training data and time constraints. The research focuses on four severely affected municipalities—Casola Valsenio, Predappio, Modigliana, and Brisighella—leveraging a unique approach where training was conducted exclusively on one municipality (Casola Valsenio) and applied to the others.
The study assesses the performance of these models across varied geological and environmental contexts, examining the impact of input data configurations, including pre- and post-event imagery, slope, and NDVI change maps derived from high-resolution aerial and Sentinel-2 satellite data. While both models achieved notable accuracy, SegFormer demonstrated greater resilience in handling complex geological conditions. Despite challenges like false positives in agricultural fields and along river margins, the models effectively reduced the time required for initial mapping, providing a critical starting point for manual refinement.
Quantitative metrics, such as F1 score and Intersection over Union (IoU), were complemented by expert qualitative evaluations, ensuring a comprehensive assessment of the models’ practical applicability. Results reveal that automated mapping, though not a replacement for manual methods, can significantly expedite the production of high-quality landslide maps, critical for immediate disaster response. By automating the initial detection and delineation processes, these methods can save weeks of work, allowing responders to focus on refining outputs and addressing urgent needs.
This research underscores the feasibility of integrating machine learning models into emergency workflows, bridging the gap between academic advancements and practical applications. Automated mapping offers a scalable, efficient, and reliable solution for rapid disaster response, particularly in large-scale emergencies, providing a foundation for future innovations in geohazard management.
How to cite: Dal Seno, N., Ciccarese, G., Evangelista, D., Piccolomini, E., and Berti, M.: Rapid Landslide Mapping During the 2023 Emilia-Romagna Disaster: Assessing Automated Approaches with Limited Training Data for Emergency Response, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6971, https://doi.org/10.5194/egusphere-egu25-6971, 2025.
Gravitational mass movements in alpine regions, such as landslides, debris flows and rockfall, are driven by complex physical processes. While the translation and runout of these events can be reasonably well modelled once they occur, the predisposing and triggering mechanisms leading to failure are very challenging to assess. This is particularly demanding for practitioners who need to take decisions on the ground to ensure the safety of the population. There is potential to improve the situation by using a variety of new space-time climate and land surface datasets to describe the hydrogeomorphic system state and relate it to possible failure by confronting it with past observed events. In this work we focus on the local susceptibility to the initiation of mass wasting events (shallow landslides, debris flows and rockfall) in low- and subalpine regions by exploring the predictive power of various hydro-meteorological drivers related to rainfall, snowmelt, high soil moisture, freezing, etc.
To provide spatially and temporally consistent information, we model all hydro-meteorological drivers governing the hydrogeomorphic catchment state of the Alpine Rhine (GR, Switzerland) over the period 1998-2022 based on globally available soil information (SoilGrids) as well as national climate (Federal Office of Meteorology and Climatology MeteoSwiss), snow (WSL Institute for Snow and Avalanche Research SLF) and terrain data (Federal Office of Topography Swisstopo). The temporal and spatial resolution of the analysis is daily over a 1x1km grid. We determine the seasonally varying contribution of each driver to the triggering of each individual mass movement type utilizing the concept of receiver operating characteristics (ROC) and its area under the curve (AUC) as performance metrics. The underlying events recorded in the Swiss natural hazard database comprise 459 shallow landslides, 295 debris flows and 761 rockfalls (StorMe, Swiss Federal Office for the Environment FOEN) in the study period. The best-performing hydro-meteorological drivers then serve as input to predict the occurrence of mass wasting events with data driven models. We test both a traditional statistical approach and machine learning algorithms to compare their capability of modelling the susceptibility to alpine mass movements.
Compared to a purely rainfall-based prediction of landslide or debris flow activity, which is commonly done in the literature, this approach benefits from the availability of further spatially distributed climate variables and terrain characteristics. Our findings contribute to a better understanding of the role of catchment state on predisposing and triggering conditions of alpine mass movements, and illustrate also the limits of predictability for such events due to the inherent randomness in the triggering processes.
How to cite: Demmel, S. and Molnar, P.: Modelling catchment susceptibility to alpine mass movements, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8086, https://doi.org/10.5194/egusphere-egu25-8086, 2025.
In recent years, the risk of landslides caused by man-made underground caves has increased on Italian territory, with significant consequences for human life and for the anthropogenic environment. Such artificial caves have generally been dug and subsequently abandoned in very soft porous rock formations, such as calcarenite deposits, even at shallow depths. The low mechanical strength values of such rocks, together with their susceptibility to weathering and consequent loss of strength, make these rock masses prone to sinkhole formation. In order to develop a rapid but mechanically based method to assess the stability of artificial caves based on the geometrical features of the cave and the mechanical properties of the rock, an improved formulation of the abaci, originally proposed by Perrotti et al. (2018), has recently been proposed by Mevoli et al. (2024), which introduces the ability to also assess the range of the cave safety factor. In this perspective, the application of the abaci can be used as a quantitative tool for the preliminary assessment of sinkhole hazards, enabling large scale analyses that can subsequently be followed by a detailed and advanced study at the local scale.
A data-driven approach was employed to compare and discuss the results obtained from the direct application of the abaci, based on this newly developed version. The selected method, proposed by Giustolisi and Savic (2006), and known as 'Evolutionary Polynomial Regression', is based on the genetic programming paradigm and returns simple functional relationships, namely polynomials of elementary functions, among the considered physical parameters. In particular, it generates a Pareto front of expressions that considers simplicity and accuracy. This facilitates the interpretation of the results of the data modelling approach, thereby maintaining focus on the physics of the phenomenon under investigation, as outlined by Fazio et al. (2024).The results will also demonstrate the use of these machine learning techniques to provide mathematical formulations that can be readily employed in the field by experts involved in assessing the stability of underground cavities.
Perrotti M., Lollino P., Fazio N.L., Pisano L., Vessia G., Parise M., Fiore A., Luisi M. (2018). Finite Element– Based stability Charts for Underground Cavities in Soft Calcarenites. Int. J. Geomechanics, 18(7), DOI: 10.1061/(ASCE)GM.1943-5622.0001175.
Mevoli, F.A., Fazio, N.L., Perrotti, M. et al. Assessing the stability of underground caves through iSUMM (innovative, straightforward, user-friendly, mechanically-based method). Geoenviron Disasters 11, 10 (2024). https://doi.org/10.1186/s40677-023-00264-3
Giustolisi O., Savic D. A. (2006). A symbolic data-driven technique based on evolutionary polynomial regression." J. of Hydroinformatics, 8 (3), 207-222.
Fazio, V., Pugno, N. M., Giustolisi, O., & Puglisi, G. (2024). Physically based machine learning for hierarchical materials. Cell Reports Physical Science, 5(2).
How to cite: Fazio, N. L., Sollecito, F., Lollino, P., and Fazio, V.: Enhancing Underground Cave Stability Assessment through Physically-Based Machine Learning Methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20095, https://doi.org/10.5194/egusphere-egu25-20095, 2025.
Landslides are one of the most serious natural disasters, causing many deaths and damage to infrastructure. In developing countries with rapidly growing cities, having accurate landslide susceptibility maps (LSM) is crucial for predicting landslides and minimizing risks. These maps play a key role in effective disaster management and mitigation strategies. While the development of advanced machine learning models such as Random Forest (RF) and XGBoost has significantly improved LSM accuracy, their complexity and "black box" nature make them challenging to interpret. This study uses SHapley Additive exPlanations (SHAP) as an explainable artificial intelligence (XAI) approach to enhance the interpretability of these ensemble models in an arid region in East Cairo, Egypt. A total of 183 landslides were identified using field surveys and satellite imagery, with 70% of the data allocated for training and 30% for validation. Fourteen predictor variables were incorporated from different categories. Both RF and XGBoost were used to create LSM, and their accuracy was compared to evaluate the most effective model. SHAP values provided a detailed evaluation of the contribution of each variable to landslide susceptibility, offering insights into the models' decision-making processes and identifying the most influential features. The results proved that SHAP not only improved the transparency of complex models but also facilitated the identification of key factors driving susceptibility, resulting in a more efficient and interpretable LSM framework. Models trained with SHAP-informed feature selection achieved high performance, with an AUC of up to 0.96. This study highlights the dual potential of explainable AI in addressing the complexity of modern machine learning models and improving their practical applicability in landslide hazard assessments.
Keywords: Landslide susceptibility, Explainable AI, Random Forest, XGBoost, Arid regions
How to cite: Abdelkader, M. and Csámer, Á.: Improving Landslide Susceptibility Mapping with Explainable AI: Enhancing Prediction and Interpretability, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7584, https://doi.org/10.5194/egusphere-egu25-7584, 2025.
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.
How to cite: Schild, L., Scheiber, T., Snook, P., Maschler, A., and Arghandeh, R.: Explainable Artificial Intelligence Based Displacement Analysis and Forecasting for Unstable Rock Slopes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17096, https://doi.org/10.5194/egusphere-egu25-17096, 2025.
Landslides are complex geohazards often driven by hydro-meteorological factors. Climate change is amplifying these drivers, potentially increasing landslide frequency and intensity. Addressing these challenges requires robust tools capable of capturing the dynamic interactions between hydro-mechanical processes. While physics-based models provide valuable insights, their reliance on simplifying assumptions limits their ability to fully represent these intricate systems. In contrast, deep learning techniques excel at uncovering non-linear interdependencies, making them well-suited for landslide modeling.
This study employs a Long Short-Term Memory (LSTM) neural network to forecast landslide displacements at the Ripley Landslide in British Columbia, Canada. Ripley is a translational landslide of significant geotechnical and environmental interest, primarily impacting major railway corridors and local river biodiversity. The landslide’s movements are influenced by a pre-sheared clay seam with residual friction angles of 9–16 degrees, as well as toe erosion and drawdown effects from the Thompson River during late spring.
Three GPS stations have monitored Ripley’s displacements since April 2008, consistently showing similar magnitudes and directions of movement. Data from one station were used to train the LSTM model, with river flow as the primary input. Synthetic noise levels were introduced into the data to evaluate model robustness, and a sensitivity analysis was conducted to examine the impact of different training datasets on displacement forecasts. Additional inputs, including temperature and precipitation, were incorporated to assess their contributions to model performance. Shapley values were employed to quantify the influence of each input variable, enhancing the explainability of the model that is typically obscured by the convoluted structure of neural networks.
This work demonstrates the potential of deep learning techniques to advance situational awareness and forecasting of landslide activity by leveraging hydro-meteorological drivers. The findings contribute to the development of data-driven approaches for landslide early warning systems and hazard mitigation strategies on a regional scale, as there are 11 other landslides in the valley within a 10-km distance that share similar surficial geology and exposure to hydro-meteorological drivers.
How to cite: Sharifi, S., Macciotta, R., and Hendry, M.: Exploring Hydro-Meteorological Drivers of Landslide Displacement: A Time-Series Forecasting Approach Using LSTM at Ripley in British Columbia, Canada, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2579, https://doi.org/10.5194/egusphere-egu25-2579, 2025.
Forecasting of landslides is crucial because these natural hazards pose significant threats to human lives, infrastructure, economies, and ecosystems. Understanding the spatial and temporal drivers of landslides enables better risk assessment, mitigation, and adaptation strategies. In previous decades, numerous studies have shown that adding hydrological information and advancements in modelling techniques have improved regional landslide early warning systems (LEWS). However, operational LEWSs are still a few. This brings up the question how the next generation LEWS needs to look like.
Landslide hazard assessment on regional scale has been founded on two main pillars: the essential inventories of slope failures and on the quantification of the hydrometeorological drivers. First, the lack of landslide inventories and the dominance of seemingly stable slopes in a region constraints our ability to empirically train landslide early warning systems. The inclusion of more multi-source slope deformation information is a logical development, however, turns out to have its own challenges; it merges different physical properties within one database. Second, causal and triggering hydrometeorological conditions are needed both in space and time for effective landslide prediction. Ideally, one would start with high spatial and temporal resolution rainfall and soil hydrological information. While acknowledging existing challenges, impressive progress has been made in this field. Combined monitoring and advanced modelling on a range of scales has resulted in valuable information on, for example, subsurface water storage. Similarly, near real-time and forecasted high resolution rainfall information from ground based rain radars shows promising results. The improved representation of the hydrometeorological conditions improves the performance of LEWS.
Starting from a brief review of the developments and limitations of regional hazard assessment, the presentation will discuss the opportunities to improve the landslide inventory site as well as through hybrid measurement and modelling approaches to quantify the dynamic hydrometeorological conditions. Landslides are an anomaly in a seemingly stable environment, and inherently, forecasting of such rare events in space and time is associated with uncertainty, but this uncertainty can be reduced which is key for protecting society from the impact of landslide hazards.
How to cite: Bogaard, T.: Challenges and opportunities in regional hydrometeorological landslide assessment, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15725, https://doi.org/10.5194/egusphere-egu25-15725, 2025.
Nowadays, the correlation between precipitation and changes in displacement rates of suspended or reactivated landslides, especially for deep-seated phenomena, is still poorly defined on a quantitative basis. This study, exploits long-term in-situ monitoring time series to propose new rainfall intensity-duration (ID) thresholds that can discriminate the acceleration of complex deep-seated landslides, including earthslides-earthflows (ES-EF), rockslides-earthslides (RS-ES), and deep-seated gravitational slope deformations-rockslides (DSGSD-RS).
The analysis focuses on 15 landslides in the Northern Apennines and Eastern Alps of Italy, which have been monitored in the period from 2001 to 2024. Monitoring was conducted using Robotic Total Stations (RTS), periodic, and continuous GNSS networks, leading to the documentation of 100 acceleration events. These events were analysed in relation to rainfall and temperature data from nearby meteorological stations, enabling the retrieval of intensity (mm/h) and duration (h) values regarding the antecedent triggering rainfall. This association was conducted considering both total rainfall (TR) and effective rainfall (ER). ER represents the amount of water potentially infiltrating in the ground having accounted for the aliquot lost due to evapotranspiration (ET) and snowfall and for the aliquot gained due to snowmelt processes.
Simultaneously, rainfall events not resulting in landslide accelerations were identified by examining the complete meteorological records for each landslide within the monitoring period. Both sets of intensity-duration records – i.e. those linked to and those independent from acceleration events – were analysed using a Receiver Operating Characteristics (ROC) approach. This method allowed to identify optimal rainfall thresholds and to compare their predictive capability with that of thresholds established by other authors for landslides occurrences.
The findings reveal that the proposed new thresholds tailored to a landslide’s accelerations dataset offer higher predictive accuracy compared to the established ones. Moreover, the study emphasizes the enhanced predictive performance achieved by incorporating effective rainfall, especially in scenarios where snowmelt contributes to landslide acceleration. These results underscore the importance of long-term in-situ monitoring and of introducing effective rainfall computations in the analysis, so to better account for various hydrological processes influencing landslide behaviour, ultimately improving early warning systems and risk management strategies for complex landslides in mountainous regions.
How to cite: Tondo, M., Mulas, M., Critelli, V., Lelli, F., Fabbiani, C., Aleotti, M., Caputo, G., Truffelli, G., Marcato, G., Mair, V., Tonidandel, D., and Corsini, A.: Long-Term In-Situ Monitoring for the Analysis of Landslides Acceleration vs Precipitation Relationships (Northern Apennines and Eastern Alps, Italy), EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11572, https://doi.org/10.5194/egusphere-egu25-11572, 2025.
Deep-seated landslides in colluvium derived from glacial sediments and shales blanket river valley slopes in the Western Canada Sedimentary Basin (WCSB) and are traversed by linear infrastructure and urban development. Porter et al (2019) estimated that the infrastructure maintenance and damage costs are in the order of $ 400 million (CDN) annually. In the spring of 2020, widespread accelerations of landslides in the northern portions of the WCSB led to the initiation of a multi-year study to better understand the relationships between short and longer-term hydroclimatic trends in relation to historical landslide activity.
Data from over 550 subsurface monitoring points (slope inclinometers and shape accelerometer arrays) were collected for over 100 slopes between the early 1980’s to present. A multi-stage cleaning process was necessary to minimize errors (installation, human, sensor) so that readings represent measurements of deep-seated landslide movement and reliably constrain discrete acceleration events. The concept of a “landslide year” was developed to delineate the annual movement cycle for landslides in the region and was defined as the period that starts in the spring when snowmelt infiltrates into the ground and finishes which the ground freezes in the autumn. Only displacement values that reliably constrained the landslide year were maintained in the database and, for sites with at least three years of readings, these values at each monitoring location were normalized against all of the readings for that site. This allowed for a more consistent comparison of the magnitude of displacements across sites and the region.
In parallel, historical hydroclimatic variables obtained from the ECWMF ERA5-Land reanalysis dataset (Muñoz-Sabater et al., 2021) were accessed, analyzed and reviewed. As with the displacement data, different approaches were assessed to provide normalized values that could represent “extreme” events and trends in the hydroclimate that could be compared across the region. The variables assessed focused on the antecedent soil moisture and the total water introduced during the landslide year from both snow melt and precipitation. These values, both absolute and normalized, allowed for both spatial and temporal analyses and data visualizations.
Random forest models were used to establish the relative importance of different hydroclimatic inputs in predicting normalized annual landslide displacements. The hydroclimatic variables seen as the most important and most useful for application in an early warning system were then evaluated in terms of their site-level “predictive power” when compared against the normalized displacement data. The test variables utilized were normalized Layer 4 soil moisture at the start of the landslide year, normalized Layer 4 soil moisture trend at the start of the landslide year and maximum normalized 60-day total water inputs within the landslide year. These tests yielded positive results in terms of correlation between combinations of the chosen hydroclimatic inputs and landslide displacement trends. Further development and testing of hydroclimate thresholds as a basis for a regional landslide awareness and early warning system is in progress.
How to cite: Froese, C., Porter, M., Sala, Z., Moorhouse, E., Coia, V., Michel, A., and Grover, P.: Relating regional acceleration events to hydroclimatic inputs for slow-moving deep-seated landslides in Western Canada, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-2183, https://doi.org/10.5194/egusphere-egu25-2183, 2025.
Global climate change creates geologic hazard cascades as the cryosphere experiences warming. The rapid retreat of Barry Glacier, a tidewater glacier in Prince William Sound, Alaska, has destabilized the cliff walls adjacent to the fjord, including a large landslide, approximately 2-km-wide, 1-km-tall, and ∼500 Mm3 in volume. The Barry Arm landslide was first identified in 2019 but has since been noted in photographs dating back to the 1930s. Catastrophic failure of the landslide has the potential to generate a tsunami with life-threatening waves in nearby fjords, including the port town of Whittier, 60 km from the landslide. Since monitoring began in 2021, slow downslope movement with short periods of acceleration has been observed. In this study, we refine the observations of landslide acceleration and correlate these periods with meteorological observations to assess the potential for further acceleration and catastrophic failure. We use ground-based synthetic aperture radar data (GBInSAR) collected sub-hourly from a location across the Barry Arm fjord since May 2022 with a line of sight that captures ~90% of the downslope landslide vector movement to generate time series of the landslide’s three main kinematic elements (distinct regions of deformation). This time series shows landslide-wide motion from late August to early November 2022 (2 months) at rates of 20-80 mm/day, then again from late September to mid-October 2023 (1.5 months) at 10-20 mm/day. No landslide-wide motion was detected in 2024. The Cascade Glacier sits stratigraphically above and to the northwest of the landslide and has been identified as a potential source of water for the landslide system. Ice-penetrating radar data collected in 2024 show an over-deepened section of Cascade Glacier adjacent to the most active kinematic element of the landslide, the Kite, suggesting melt water might pool and subsequently seep into the Kite kinematic element. Two full meteorological stations, each with additional node stations, monitor weather near the landslide and provide 15-minute precipitation and temperature data. We combine a simple positive degree-day factor melt model with precipitation analysis to show that the timing of movement of the Kite is correlated with the effects of seepage into the landslide subsurface, which are primarily driven by snow and ice melt. Understanding links between landslide displacement and melting of snow and ice could potentially lead to the use of meteorological conditions or forecasts as an additional risk assessment tool for identifying when the hazard of failure could be most severe. Our study accompanies others’ analyses of the Barry Arm Landslide using lidar, satellite InSAR, seismic, and infrasound data and contributes to our limited but critical understanding of landslide hazards in Alaska.
How to cite: Dow, H., Collins, B., Wolken, G., Miles, C., and Gassner, J.: Meteorological drivers of seasonal motion at the Barry Arm Landslide, Prince William Sound, Alaska, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14690, https://doi.org/10.5194/egusphere-egu25-14690, 2025.
Flash floods frequently co-occur with landslides, during which landslides can deliver large amounts of hillslope material into the river system. Their interaction can lead to exacerbated and destructive impacts. While such geo-hydrological hazards are typically triggered by intense rainfall over only a few hours, daily to monthly variations in rainfall drive soil moisture changes and alter their likelihood of occurrence, alone or in combination. The influence of this preconditioning rainfall on compounding landslides and flash floods, however, remains overlooked. Acquired through the combined use of optical and radar satellite imagery, we present a unique multi-temporal inventory of a hundred new landslide and flash flood events located in a large region in the African tropics that is characterized by active rifting and strong human influences on the landscape. From this inventory we show that preconditioning rainfall plays a central role in the occurrence of landslide and flash flood events, along with land use/land cover and landscape geological history. Wetter-than-average conditions in human-dominated cultivated areas on rejuvenated hillslopes associated with the rift formation more frequently lead to compounding flash floods and landslides. On the other hand, drier-than-average conditions in forested regions outside these rejuvenated landscapes more often lead to compounding, densely spaced and larger landslides without flash floods. This research shows that preconditioning rainfall can exacerbate the severity of co-occurring and interacting landslide and flash flood events, stressing the need to understand these geo-hydrological hazard in a compounding manner.
How to cite: Deijns, A., Thiery, W., Déprez, A., Dille, A., Malet, J.-P., Maki Mateso, J.-C., Michéa, D., Mugisho Bachinyaga, J., Sekajugo, J., Sibomana, P., Zscheischler, J., Kervyn, F., and Dewitte, O.: How preconditioning rainfall controls landslide and flash flood events in tropical East Africa, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10781, https://doi.org/10.5194/egusphere-egu25-10781, 2025.
The Himalayan region is prone to numerous landslides, primarily triggered by heavy precipitation. Most of these landslides occur from June to September, coinciding with the monsoon period. Therefore, monitoring rainfall intensity is vital for landslide risk assessment in the Himalayas. However, the sparse network of rain gauges in this region poses a significant challenge for climate extremes research. Satellite and Land Surface Model-derived precipitation products can help assess climate risks like landslides and floods without the need for installing rain gauges in remote locations. This study compares gauge-based and satellite-based precipitation products at 25 different locations using various statistical tools to evaluate their performance in landslide hazard assessment in the Himalayas. Based on statistical metrics, the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) demonstrated the highest efficiency in reproducing spatiotemporal precipitation patterns at landslide-prone sites. The comparison involved metrics such as Root Mean Square Error (RMSE), Pearson Correlation Coefficient (CC), and Relative Bias (RB), along with rainfall frequency indicators and intensity comparisons. ERA5 emerged as the best-performing product, with RMSE ranging from 2.31 to 29.80, the highest CC, and the minimum RB at most sites. It successfully estimated 5761 days of very heavy rainy days (>20mm) compared to 5014 days recorded by rain gauges. Additionally, the correlation for rainfall intensity over a 30-day cumulative period was highest for ERA5 at most sites. The role of antecedent soil moisture in triggering of landslides cannot be ignored. However, in situ soil moisture data are rarely available in hazardous zones. The advanced remote sensing technology could provide useful soil moisture information. The study explores the use of GLDAS soil moisture product at the root zone depth along with ERA5 precipitation over a prolonged period to calculate thresholds for landslide initiation under different environmental conditions over the Indian Himalayas. The study reveals that certain combinations of Land Use Land Cover classes and soil types, especially on steeper slopes, are more susceptible to landslides, with landslides being triggered even at relatively low levels of soil moisture and precipitation.
How to cite: Sharma, S., Barbhuiya, S., Gupta, V., and Das, S.: Comparative Analysis of Satellite and Gauge-Based Precipitation Data for Landslide Risk Assessment in Himalayas, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19794, https://doi.org/10.5194/egusphere-egu25-19794, 2025.
To reduce the impacts of debris flows, debris avalanches and slushflows the Norwegian Water Resources and Energy Directorate (NVE) is operating a forecasting and early warning service that issues daily landslide warnings to local authorities and public in general. Already in the first 5 years of operations, it has been observed that the most relevant landslide-triggered hydro-meteorological conditions (LHMC) vary between regions and seasons. Two different approaches have been tested to further explore this observation.
Using a heuristic approach, based on observations, region and season specific LHMC have been identified. These conditions are defined by the spatial and temporal distribution of different hydro-meteorological parameters (e.g. rainfall, snowmelt, soil saturation, etc.), landslide occurrence, as well as other synoptic conditions (i.e. information about location and paths of low- and high-pressure systems, coincidence of atmospheric rivers, strong wind, extreme events, etc.). The landslide data are obtained from the national mass movements database available at www.skredregistrering.no, while historical hydro-meteorological data are recorded as 1km2 grid maps at seNorge.no.
The analysis confirmed that water, in form of rainfall (also convective), snowmelt, high soil saturation or a combination of them, is the main triggering mechanism of landslides. In total eight hydro-meteorological conditions have been found to be most relevant for landslide occurrence. Each LHMC is described based on certain criteria like: main exposed areas, temporal distribution (season and month), general weather description and type of weather prognosis, duration of the condition, other synoptic information, list of dates when the condition was observed and caused landslides, general description of the main hydro-meteorological parameters, number and type of landslides, information about other associated hazards, evaluation of the landslide hazard index performance and recommendation about the most appropiate warning level.
Separately, a quantitatively evaluation was also tested, in a selected region, by using rain as main triggering factor, and the Grosswetterlagen (GWL) weather pattern classification through exploratory and statistical analysis, to see how this can be used as integrated tool in the operational service.
In this work, the applied analytical process is described. The hydro-meteorological conditions and their predictability are also shortly described, by presenting some recent examples. Finally, it is explained how the LHMC are integrated in the daily forecasting operations. Ideas for improvements will be discussed.
How to cite: Devoli, G., Fleig, A. K., and Olvin, V. K. M.: Preliminary identification of hydro-meteorological conditions that trigger landslides in Norway, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12687, https://doi.org/10.5194/egusphere-egu25-12687, 2025.
Rainfall-induced shallow landslides are expected to change in frequency and distribution as a result of altered patterns and intensity of rainfall. Yet, linking climate change effects to past occurrences is challenging due to the lack of long-term, systematic, and reliable datasets of landslide events. However, the widely observed increase in the number of recorded landslides over time may also be indicative in the extent of exposed assets and their vulnerability, as well as the more comprehensive event documentation carried out in recent years, rather than reflecting the actual impacts of climate change.
To decipher such a conundrum, a high-resolution space-time data-driven model recently developed and trained for well-observed time periods within the territory of South Tyrol (Italian Alps) was used to create a continuous dataset of daily landslide hindcasts (i.e. modelled probabilities) to be used as a proxy for critical conditions of landslide occurrence in space and time. High landslide probabilities in the dataset can be linked to recorded landslides, but could also represent nearly-missed events, landslides that occurred but were not recorded (for example, those that happened in remote areas away from infrastructures), or to model errors.
Daily landslide probability predictions were obtained on a 30mx30m grid for the years 1980-2020, using both static (topography, geologicy and vegetation) and dynamic factors (antecedent and triggering precipitation, and seasonal effects). The results were aggregated over 5261 slope units identified for South Tyrol, which better reflect the hydrological and geomorphological processes shaping the landscape providing, at the same time, consistent geographical boundaries to manage the aleatory uncertainty of the model.
This new enriched dataset has been used to explore changing trends and patterns in landslide probability predictions and investigate underlying causes, such as the role of the Jenkinson and Collison weather types in shaping the spatial patterns of probability predictions.
Our results could improve the ability to predict critical conditions for landslide occurrences in the future, thereby offering new tools for mitigation and adaptation strategies, and specifically supporting the elaboration of efficient early warning systems.
How to cite: Zennaro, B., Zebisch, M., Pittore, M., Lemus i Cànovas, M., Comiti, F., and Steger, S.: Deciphering landslide occurrence under climate change in South Tyrol (Italian Alps) using interpretable data-driven models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16906, https://doi.org/10.5194/egusphere-egu25-16906, 2025.
European weather services are currently transitioning from traditional weather warnings to impact-based warnings (i.e., from "what the weather will be" to "what it will do"). To inform on what impacts can be expected, meteorological data must be integrated with data on potential hazards and elements at risk.
In this study, we developed three impact models on a daily scale to predict the impact of mass movement across the entire Alpine region (450,000 km²). The models focused on three major process classes (slide-types, flow-types, and fall-types) that impact infrastructure, such as buildings and roads. The study area was first divided into ~18,000 sub-basins, with potential process areas (PPAs) delineated in each basin using the angle of reach principle and random walk routing. PPAs enabled a tailored preparation of data describing environmental drivers (e.g., morphometry, land cover, lithology), dynamic meteorological data (e.g., antecedent precipitation, short-term precipitation, temperature effects), and exposure (e.g., number/density of buildings/roads within the PPA). The impact data consisted of precipitation-induced mass movements in Austria and northern Italy, covering more than 3600 basins. This training area was considered sufficiently representative of diverse Alpine environmental conditions to allow for spatial model transferability. Additional steps involving data sampling and the reclassification of predictor variables further supported the extension of model predictions beyond the training area. For example, lithology and land cover data was reclassified to ensure that each unit within the Alpine Space was adequately represented in the training data.
Generalized additive mixed models (GAMMs) with automated variable selection were used to link binary impact data to driving factors. Rigorous evaluations, including cross-validation and feature importance assessments, showed high predictive performance (e.g., AUROCs > 0.8) and plausible relationships between drivers and impacts. For example, impact probabilities for slide-types were modeled to be highest when intense short-term precipitation followed high antecedent rainfall, particularly in drier regions that are less "adapted" to such events. Further, a higher number/density of buildings or roads within PPAs also increased impact likelihood, while effects related to morphology, temperature, lithology, land cover, and seasonality further supported model plausibility. The applicability of the model is presented from three perspectives: (i) "What-if" scenarios to explore how hypothetical changes in drivers (e.g., precipitation) affect impact probabilities; (ii) hindcasting to validate model predictions for past events and demonstrate potential for impact-based early warning; and (iii) trend analysis, using ~6,000 daily hindcasts (2005–2021) to reveal spatio-temporal trends through the lens of climate change.
The research leading to these results has received funding from Interreg Alpine Space Program 2021-27 under the project number ASP0100101, “How to adapt to changing weather eXtremes and associated compound and cascading RISKs in the context of Climate Change” (X-RISK-CC).
How to cite: Steger, S., Spiekermann, R., Lehner, S., Enigl, K., Moreno, M., Crespi, A., and Schlögl, M.: Data-driven modeling of mass movement damage potential across the Alpine Space: A step toward impact-based early warning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9819, https://doi.org/10.5194/egusphere-egu25-9819, 2025.
Posters on site: Fri, 2 May, 14:00–15:45 | Hall X3
Landslide is one of the most common natural hazards in Taiwan. Because of the complicated terrain, geological, geotechnical and weather conditions in Taiwan, landslides are frequently triggered by earthquakes, typhoons or heavy rainfalls almost year-round, posing significant threats to human lives and property and sometimes causing catastrophic damages. Rapid and accurate detection and classification of landslides are crucial for disaster mitigation, management and prevention. In this regards, satellite remote sensing is an effective approach for collecting data. However, accurate mapping and monitoring landslides usually requires analyzing considerable amounts of images, which is time-consuming and labor-intensive. In addition, in some mountainous regions, landslides may occur repeatedly, and old landslides affected areas may be reclaimed by vegetation, making it difficult to fully understand the spatio-temporal characteristics and changes of landslides. To address these issues, this study adopts a deep learning framework, TransUNet, and develops a two-stage training process and data stacking strategy to detect and classify landslide changes from multi-temporal satellite images of a mountainous watershed region is southern Taiwan. TransUNet combines the strengths of Convolutional Neural Networks (CNNs) and Transformers. Three benchmark datasets (Landslide4Sense, HR-GLDD, and Bijie Dataset) were evaluated in conjunction with labelled image titles extracted from collected SPOT satellite images of the study area for transfer learning. Training of the deep learning model was separated into two stages: the first stage focused on initial landslide change detection, while the second stage refined the classifications by applying a weighting scheme. Results of this study show that TransUNet performs well with high-resolution satellite images for landslide change detection, with the best Precision, Recall and F1-Score of 0.92, 0.76 and 0.82, respectively. In addition, despite lacking a temporal feature extraction framework, developed model can effectively distinguishes the changes of landslide affected areas such as old landslides, new landslides, and vegetation reclaimed areas.
How to cite: Tsai, F. and Tsai, S.-N.: Landslide Change Detection from Satellite Images with Deep Learning Classification, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12109, https://doi.org/10.5194/egusphere-egu25-12109, 2025.
Slow mass movements are widespread players of slope dynamics, with different mechanisms depending on involved materials and geomorphic settings. Alpine para/periglacial environments are extensively affected by slow rock-slope deformations, deep-seated rock and debris slides, and active periglacial features, while fluvial-dominated mountain ranges are typically affected by rapid rockslides and long-lived earthflows. These processes exhibit different deformation patterns and rates, threatening lives and infrastructures in different ways. Mapping and monitoring slow mass movements is thus essential for civil protection, land management, and disaster risk reduction, requiring capabilities to rapidly map and classify processes over large areas.
Current regional-scale approaches to capture mass movement activity rely on geomorphological techniques supported by remote sensing. These approaches are accurate but time consuming and difficult to update. Such gaps could be filled using artificial intelligence techniques, currently mostly based on the interpretation of optical imagery or multitemporal InSAR data. Nevertheless, mass movements are often too fast to be captured by multitemporal InSAR and too slow for optical or amplitude SAR image analysis. Dual-pass satellite DInSAR products offer a valuable alternative to study these intermediate processes by the analyses of interferometric fringes, yet they suffer from noise, artifacts, and unwanted signals due to atmospheric disturbances.
We propose a deep learning model to automate the detection and classification of different types of mass movements in different geological and geomorphological settings through the interpretation of deformation fringes in DInSAR interferograms. To this aim, we use a YOLO, a convolutional object detector, aimed at interpreting routinely available wrapped interferograms. To mirror the interpretative process carried out by a human expert, input data include interferograms, a compound measure of the reliability of the interferogram, and a composite layer of geomorphological and morphometric information.
To train our net, we developed a geomorphologically constrained methodology to construct libraries of labeled expert-interpreted InSAR phase signal, corresponding to different mass movements recognized in two large (103 km2) test areas in the Central Alps (Lombardia) and Apennine (Emilia-Romagna) of Italy, representing diverse processes and geological settings. The model is tested with sets of routinely generated SAR interferograms, to produce automated maps able to detect and classify mass movements over different timescales. This approach promises to streamline the rapid generation and update of active landslide inventories, to support local-scale landslide monitoring plans and civil protection actions, and improve the integration of data into landslide modeling efforts.
How to cite: Mondini, A., Bovenga, F., Simoni, A., Reyes-Carmona, C., Mercurio, A., and Agliardi, F.: Automated detection of active mass movements in SAR interferograms using Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16692, https://doi.org/10.5194/egusphere-egu25-16692, 2025.
Landslide inventories are essential for hazard assessment and risk mitigation, yet their accurate and efficient creation remains a challenge, particularly in forested and topographically complex regions. Traditional approaches relying on RGB imagery often struggle with dense vegetation cover, which obscures landslide features. In this study, we propose an innovative deep learning framework utilizing the Segment Anything Model with Low-Rank Adaptation (SAMLoRA) to automatically detect and map landslides from hillshade datasets. Hillshade representations, derived from high-resolution Digital Elevation Models (DEMs), provide enhanced visibility of topographic features by emphasizing surface morphology independent of vegetation cover.
Our model was trained on a diverse dataset collected from Romania, Czechia, and Italy, comprising over 5,000 manually delineated landslide polygons. By leveraging the SAMLoRA model, which combines the robust segmentation capabilities of SAM with the adaptability of LoRA for domain-specific fine-tuning, we achieve superior landslide detection performance compared to RGB-based methods. Our approach effectively identifies landslides even in densely forested areas, where traditional image-based techniques often fail. Experimental results demonstrate that the SAMLoRA model achieves an accuracy exceeding 80%, significantly improving both precision and recall while reducing manual mapping efforts.
This study highlights the potential of deep learning applied to topographic derivatives, paving the way for more reliable and automated landslide inventory mapping in diverse and challenging environments.
How to cite: Sandric, I., Ilinca, V., Letal, A., Raj Meena, S., Irimia, R., Botea, A., Catani, F., Chitu, Z., and Klimes, J.: Automated Landslide Inventory Mapping Using SAMLoRA and Hillshade Datasets: A Deep Learning Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18812, https://doi.org/10.5194/egusphere-egu25-18812, 2025.
The Himalayan terrain has encountered multiple vandalized events that have hampered humans and property. While significant progress has been made in leveraging Earth Observation data for landslide mapping, several critical challenges remain in creating models that can be operational globally. The first limitation is that no high-resolution, globally distributed, and event-diverse dataset is available for landslide segmentation. Inadequacy in data impairs the ability of machine learning models to achieve accurate and robust detection over different terrains since insufficient representation of both landslide and non-landslide classes leads to suboptimal generalization. We provide the High-Resolution Global Landslide Detector Database (HR-GLDD) to fill this critical gap. The unprecedented dataset, derived from PlanetScope imagery with an extraordinary 3-meter pixel resolution, includes a detailed set of landslide instances, including those from the Kalimpong Himalayas in Northeast India, providing never-before-attempted granularity and diversity for global landslide modeling.
The HR-GLDD contains ten independent landslide events, five rainfall-triggered and five seismic, under diverse geomorphological and topographical conditions. Standardized image patches from high-resolution PlanetScope optical satellite imagery in four-spectral-band (red, green, blue, near-infrared) combinations of bands and binary masks delineating landslides are provided. One of the first datasets prepared for landslide research using high-resolution images in artificial intelligence for landslide detection and identification studies is particularly relevant using HR-GLDD.
Five state-of-the-art deep learning models were utilized to validate its usefulness by showing stable performance at Kalimpong, verifying the dataset's robustness and transferability. HR-GLDD is publicly available and valuable for calibrating and building models to produce reliable landslide inventories after an event. The constant updating of data from recent landslide events significantly increases its usefulness in developing landslide research and risk assessment.
How to cite: Singh, S., Raju, A., and Meena, S. R.: Adaptive Deep Learning Framework for Rapid Landslide Mapping Using HR-GLDD, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1348, https://doi.org/10.5194/egusphere-egu25-1348, 2025.
The volcanic area in Indonesia is geologically characterized by the presence of pyroclastic products, which are prone to intense weathering and thus susceptible to different types of landslides. Combined with adverse weather conditions related to the tropic climate, landslide activity is generally high in volcanic soils, leading in the past to numerous events in the Cipongkor District, West Bandung, in Indonesia. On March 28th, 2024, a landslide affected a densely populated settlement area, destroying some houses and impacting the provincial road crossing the landslide area.
This research investigates the geological, geomechanical and hydrogeological characteristics of the slides and proves the influence of precipitation on the initial formation process. The applied methods are manifold and comprise UAV-based aerial mapping supported by geomorphological-geological field observations, geotechnical drilling including core sampling, geomechanical properties examination, analyses of meteorological data, and numerical modeling. The geometry and volume of the landslide were determined by UAV and field mapping by reconstructing the pre-failure topography. The lithostratigraphic data obtained from the borehole are improved by resistivity (ERT) measurements, in order to build a geological subsurface model of the slide. Based on this and considering hydrogeological and geomechanical data numerical modeling is applied to simulate the initiation of the slide by applying a transient approach which is able to study precipitation data, pore pressure changes and slope failure.
Preliminary results show that the stratification of ash tuff and lapilli layers, with their variation of weathering may provide a disposition factor for the formation of the slide. Data from the nearest local meteorological station show that cumulative precipitation in the research area during the rainy season (October 2023 to March 2024) was 1230 mm. Furthermore, in the 3 consecutive days before the slide event precipitation reached 95 mm, suggesting that heavy precipitation may have acted as a trigger that caused the failure event of this first-time slide.
How to cite: Misbahudin, M. and Zangerl, C.: Investigation and Characterization of Landslides in Volcanic Soils Triggered by Rainfall in West Bandung, Indonesia, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4613, https://doi.org/10.5194/egusphere-egu25-4613, 2025.
Landslides pose significant risks, often causing severe property damage and, in extreme cases, loss of life due to poorly timed evacuations. Accurate forecasting is, therefore, essential. Traditional landslide studies rely heavily on satellite imagery to analyze timing and impact, often using machine learning models to process these images or predict landslides based on relevant factors. However, the lack of sufficient data significantly compromises forecasting accuracy in data-scarce regions such as remote mountainous areas or highways. Federated learning, a cutting-edge machine learning paradigm, offers a promising solution by aggregating model parameters from decentralized edge models operating in different regions. This approach allows a central model to leverage diverse, region-specific data without requiring direct data sharing, resulting in a more robust and generalized predictive capability. The framework supports edge models that process localized data varying in both temporal and volumetric dimensions, while a carefully designed parameter aggregation mechanism ensures iterative improvement of the central model. Experimental results demonstrate that federated learning enhances forecasting performance and improves accuracy, particularly in regions with limited data availability, marking a significant step forward in landslide forecasting.
How to cite: Cheng, P.-W. and Tsai, W.-P.: Federated Learning-Based Approach for Landslide Forecasting in Taiwan, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6125, https://doi.org/10.5194/egusphere-egu25-6125, 2025.
Deep-seated landslides in overconsolidated glaciolacustrine materials typically cycle through episodic periods of gradual acceleration and deceleration. The 1973 Attachie landslide (BC, Canada) and 2014 Oso landslide (Washington, USA) are well-known examples of landslides that deviate from this trend, instead failing extremely rapidly, with considerable runout that dammed the Peace River (Attachie) and impacted a community nearly 1.5 km away, resulting in 43 fatalities (Oso). Given the velocity and runout distance of these two landslides, further characterization of the landslide, priming, and trigger mechanisms may help manage geohazard risk for other landslides in similar terrain.
The landslide mechanisms and antecedent climatic conditions prior to failure have been relatively well studied for the Attachie and Oso landslides. As part of these studies, hydroclimatic re-analysis tools have been applied, correlating soil moisture data with precipitation records to understand the dominant timescale by which hydroclimatic conditions may have triggered activity within these landslides in the past.
In spring 2022, another extremely rapid landslide derived from glaciolacustrine materials occurred on the Halfway River, less than 10 km away from and initiating within the same geological unit as the 1973 Attachie landslide. The objectives of this study are twofold: to apply the same hydroclimatic re-analysis and precipitation review methodology to the Halfway River landslide, and to compare hydroclimatic trends across all three landslides. Comparison of landslide morphology, mechanisms, and material properties between these landslides are left to future research.
Soil moisture and precipitation data were obtained from the land component of the ERA5 climatological re-analysis data produced by Copernicus Climate Change Service of the European Union. At the Halfway River slide, soil moisture (1-3 m depth) was above the monthly average for 65% of the months since over the 8-year period prior to the failure, with above-average annual soil moisture in 5 of the 8 years. Soil moisture and precipitation at the time of failure were not exceptional, although the failure occurred during the first rain-on-snow event in above-zero °C conditions of the year, which may be the triggering event. Annual precipitation and soil moisture in the year prior to the April 2022 failure were below average, indicating that one year of drier-than-average conditions may be insufficient in arresting the deformation processes that are hypothesized to predicate these extremely rapid failures.
No discrete trigger was identified for the Attachie landslide. The dominant theory is that a longer-term internal deformation and acceleration trend associated with a 10-to-15-year period of above-average soil moisture preceding the 1973 failure caused the event. At the Oso landslide, a possible triggering event was identified from a nearly one in 10-year soil moisture peak, resulting from both a longer-term elevated soil moisture trend and three weeks of intense rainfall. This occurred in the context of a 4-year period of above-average precipitation. While it is likely that a variety of processes contributed to the extremely rapid failures of these landslides, these examples support the current hypothesis that multi-year moisture trends drive gradual deformation, preconditioning these slopes for extremely rapid failures.
How to cite: Funk, A., Tauskela, L., van Veen, M., Mitchell, A., and Porter, M.: An evaluation and comparison of hydroclimatic data preceding extremely rapid glaciolacustrine landslides, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13845, https://doi.org/10.5194/egusphere-egu25-13845, 2025.
The Upper Ormonts Valley (Ormonts-Dessus), located in western Switzerland, corresponds to the catchment area of the Grande Eau River and is located on the border between the Pre-Alps and the Alps. The valley has a general east-west orientation and is bounded by the Pic Chaussy – La Para massif to the north, the Diablerets massif to the east and southeast, and the Chamossaire – Col de la Croix massif to the south. Historically, it has been exposed to many natural hazards such as avalanches, floods, landslides, rockfalls and debris flows. The southern slope of the Pic Chaussy – La Para massif, facing the valley, is subject to avalanches as well as rockfalls, debris flows and shallow landslides. This slope has been monitored using temperature sensors near the summit, combined with data from a SLF weather station (Swiss National Institute for Snow and Avalanche Research), and annual lidar scans from the opposite side of the valley. In the Diablerets massif, two tributaries of the Grande Eau River, the Dar (10 km2) and the "upper" Grande Eau (12 km2), were also studied. After the confluence of the two alpine streams, the Grande Eau flows through the village of Les Diablerets, a major tourist destination in the area. Here, floods and high bedload events have occurred, and riverbank erosion is common. The Dar glacial cirque is an area of high sediment production due to permafrost thaw, while landslides are common in the lower part of the Dar catchment. Both tributaries have been monitored using time-lapse wildlife cameras and annual lidar scans. The Dar catchment has been studied more extensively using DoD’s, drone orthomosaics, lidar scans and sediment budget estimates. A drone lidar scan is planned for this spring. Despite the short observation period (2023-2024), some drivers of change have been identified. Mild winters and wet springs such as that of 2023/2024 resulted in exceptional precipitations at mid-elevations, as well as large daily temperature variations at high elevations. Wet conditions such as these favored shallow landslides, strong riverbank erosion and a few high discharge events in the Grande Eau River. Changes in rockfall frequency have not yet been observed. And the effects of a stronger winter like 2024/2025 remain to be seen.
How to cite: Gutierrez, A., Derron, M.-H., Gerber, C., Gendre, N., Werren, G., and Jaboyedoff, M.: Monitoring current impacts of climate change on slope stability in the Ormonts valley, western Switzerland, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15564, https://doi.org/10.5194/egusphere-egu25-15564, 2025.
Taiwan, situated at the junction of the Ryukyu Arc and the Philippine Arc, is prone to frequent seismic activities due to its position at the boundary of tectonic plates. Earthquake-induced landslides, therefore, are one of the most common geological hazards. For disaster mitigation, it is crucial to accurately predict the spatial distribution of such landslides after earthquake occurrence. This study revolves around assessing the landslide risks triggered by the April 3rd, 2024, Hualien earthquake, which caused tremendous damage and claimed 18 lives, using multiple machine learning models, including Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting Machine (GBM), and K-Nearest Neighbors (KNN). However, Logistic Regression (LR) was undiscussed in this study due to its disaster prediction limitations. While LR is advantageous when handling small datasets with limited independent variables, it faces significant drawbacks in high-dimensional and multi-variable scenarios. Moreover, the simplistic structure of LR tends to result in underfitting, causing inferior predictive performance. Furthermore, when dealing with large-scale data, the process becomes computationally intensive for LR. In contrast, machine learning models like RF, SVM, and GBM, along with ensemble techniques, are better suited for addressing the complexity of earthquake-induced landslide prediction.
The models were trained using a dataset comprising 3191 data points, including various topographic, geological, and seismic variables such as slope-related factors, curvature, elevation, aspect, lithology, peak ground acceleration (PGA), peak ground velocity (PGV), and distances to nearby faults and rivers. The dataset was labeled into two categories: coseismic landslide (CL) data labeled as 1 and non-coseismic landslide (NCL) data labeled as 0. To train and evaluate the models, the dataset was divided into two subsets: 70% was used as the training set to build and fine-tune the models, while the remaining served as the test set to assess their predictive performance. The confusion matrices of the four models were the basis for comparing their performance. All models’ accuracy exceeds 0.95. Among them, the SVM model reached the highest at 0.9822, followed by GBM (0.9702), RF (0.9697), and KNN (0.9530). The greater performance of SVM can be attributed to its ability to handle high-dimensional and nonlinear data more effectively, using kernel functions to transform the feature space and maximize the margin between classes, enhancing its classification precision and generalization capability.
To further enhance prediction reliability, an ensemble model was developed by integrating the RF, SVM, and GBM models, while the KNN model, showing the lowest accuracy, was excluded, ensuring the number of the models was odd. The final prediction of the ensemble model was voted by the outcome of the three models, substantially reducing prediction errors.
Compared to logistic regression models, the ensemble approach is more dependable. While logistic regression struggles with high-dimensional, non-linear, and strongly correlated geophysical variables, the ensemble model formed by three machine learning models (RF, SVM, and GBM) combines their strengths to tackle these challenges. By leveraging the models’ diversity, the ensemble reduces overfitting and enhances the robustness of predictions, highlighting the ensemble model’s capability in addressing the complexities of coseismic landslide prediction.
How to cite: Ou Yang, Y. H., Chao, W. A., and Yang, C. M.: Machine Learning for High-Accuracy Co-Seismic Landslide Risk Prediction Using Multi-Parametric Data: A Case Study of M7.2 Hualien Earthquake, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10556, https://doi.org/10.5194/egusphere-egu25-10556, 2025.
