Hybrid, physically constrained machine learning models combine the predictive power of machine learning approaches with the plausibility and interpretability of established physical models. The architecture of artificial neural networks (ANNs) allows to incorporate process-based constraints and physical laws to ensure a physically plausible and therefore generalizable model output.
Hybrid models have proven their utility in a variety of scientific domains and, most recently, in the Earth system sciences. They have been successfully applied to model the global hydrological cycle or ocean currents and sea surface temperatures.
However, up to now, the applicability of hybrid models has not yet been explored for landslide susceptibility and hazard modeling.
It is therefore our objective to shed light on the potential of hybrid, physically constrained slope stability models by assessing the predictive performance and plausibility of results as a prerequisite for a wider adoption of such approaches in landslide studies. We have embedded an established slope stability model in an ANN framework to overcome parameterization issues: The ANNs estimate the spatial distribution of soil properties and local soil cohesion as spatially variable latent inputs to the physically based model structure without requiring field or laboratory data of these parameters. As a case study, in cooperation with the Geological Survey of Slovenia (GeoZS) we have developed a landslide susceptibility map for the municipalities most affected by the disastrous rainfall event in August 2023.
Preliminary results show a good agreement with existing susceptibility maps produced with traditional slope stability models. Model parameters which would require extensive laboratory measurements for calibration could be plausibly estimated by machine learning. The hybrid approach furthermore allowed us to explicitly map these latent variables as a side product that supports model interpretation and can be evaluated with ancillary data that may become available in the future.
Building upon these results, we plan to expand the model's spatial and temporal domains. In doing so, we can assess this novel approach in terms of its transferability and generalization capabilities.

How to cite: Strohmaier, F. and Brenning, A.: Hybrid Physically Constrained Machine Learning Models of Landslide Susceptibility: a Case Study from Slovenia, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17248, https://doi.org/10.5194/egusphere-egu24-17248, 2024.

The Emilia-Romagna Region in Italy faces significant challenges due to landslide hazards. With over 80,000 landslides identified in its mountainous regions, some areas see more than a quarter of their land impacted. Despite the generally slow nature of these landslides, they pose a considerable economic burden. For instance, in 2019, the region allocated 1 million euros for immediate safety measures, and it's estimated that an additional 80 million euros are needed to complete safety plans. This makes Emilia-Romagna one of the most landslide-prone areas globally. Factors like the region's geological makeup, increased land use, and climate change are exacerbating the issue. It's becoming evident that emergency measures alone are insufficient, and proactive prevention strategies are essential. Key efforts include better forecasting of rain-induced slope instabilities and predicting reactivations of dormant landslides and new failures. However, the unpredictable nature of landslides makes these goals challenging.

The primary aim of this study is to create AI models to predict landslides in Emilia-Romagna, leveraging 75 years of data collected by the University of Bologna in partnership with the Regional Agency for Civil Protection and the Geological Survey of Emilia-Romagna. Various methods like Bayesian analysis, Neural Networks, XGBoost, TPOT, Random Forest, LDA, QDA, and Linear Regression have been employed. The findings suggest that landslides in this region are primarily driven by rainfall during the event and its location, while prior rainfall seems less critical. The research also found that after a dry summer, a rainfall event of 90-100 mm is typically needed to trigger a landslide, a threshold that decreases later in the year. The best algorithm had an F2 score test result of 0.6, meaning it could correctly predict a true positive (rainfall causing landslide) every 3 positive instances and correctly predict a true negative (rainfall not causing landslide) 95.5% of the time.

How to cite: Dal Seno, N., Evangelista, D., Piccolomini, E., and Berti, M.: Comparative analysis of conventional and machine learning techniques for rainfall threshold evaluation under complex geological conditions, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10837, https://doi.org/10.5194/egusphere-egu24-10837, 2024.

The presence of vegetation plays an important role in slope stability, especially in triggering of shallow landslides. It influences the mechanical and hydrological behaviour of soils, generating both stabilizing and destabilizing actions [1,2]. Variation in vegetation related to land use change can affect slope stability and can be evidenced in terms of variation of probability of failure.
In this study we implement a module for the calculation of root reinforcement in the slope stability physically-based probabilistic model PG_TRIGRS (Probabilistic, Geostatistic-based, TranSient Rainfall Infiltration and Grid-based Slope stability, [3]). Such model allows the wide-area assessment of the probability of rainfall-induced failure, considering the spatial variability of the soil properties treated as random variables. In this work, we apply the model to an area prone to landslides in Central Italy assuming the spatial variability of vegetation.
To investigate the influence of the spatial layout of plant roots on slope stability, the root reinforcement is implemented in the PG_TRIGRS probabilistic model. The considered root cohesion values  were derived from literature and were determined according to vegetation maps available for the study area. In addition, root cohesion variation is also considered along the vertical profile as a function of rooting depth. Finally, the resulting probability of failure distribution is compared to the results obtained for the bare soil with the absence of roots.

[1] Pollen-Bankhead, N., & Simon, A. (2010). Hydrologic and hydraulic effects of riparian root networks on streambank stability: Is mechanical root-reinforcement the whole story?. Geomorphology, 116(3-4), 353-362.
[2] Masi, E. B., Segoni, S., & Tofani, V. (2021). Root reinforcement in slope stability models: a review. Geosciences, 11(5), 212.
[3] Salciarini, D., Fanelli, G., & Tamagnini, C. (2017). A probabilistic model for rainfall—induced shallow landslide prediction at the regional scale. Landslides, 14, 1731-1746.

How to cite: Galeazzi, S., Ciabatta, L., Brocca, L., and Salciarini, D.: A probabilistic model for slope stability analysis including the root reinforcement effects, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-5565, https://doi.org/10.5194/egusphere-egu24-5565, 2024.

Dams, powerlines, and power plants represent strategic energetic infrastructures and their future operativity maintenance is a challenge. Stakeholders are strongly interested in evaluating the potential risks that may affect their functionality, especially regarding natural hazards. In Italy, geo-hydrological hazards triggered by rainfall such as floods and landslides represent a serious threat to electrical infrastructure, since their magnitude is generally difficult to modelling and quantify properly.

Here, we present an application of the model proposed by Borga et. Al. for rainfall-induced shallow landslide hazard assessment. The model merges an infinite slope stability equation with a simplified hydrogeological model evaluating, for a defined rainfall duration, the critical rainfall ratio able to trigger the landslide failure. The model has been adapted to work automatically using Python scripts and has been extended proposing a new strategy for evaluating the Dynamic Contributing Area and for including soil moisture information. Rainfall return time was considered as a proxy of the magnitude of the geo-hydrological events, identifying the most hazardous area with respect to the position of powerlines for the case study basin of Trebbia River, Emilia, Italy. Model results were validated against the currently available local rainfall threshold curves, showing good skill in failure detection.

The instrument could be useful for planning purposes, addressing, and quantifying the location under which the critical infrastructure may encounter risk with respect to geo-hydrological threats, and giving useful insights about possible mitigation strategies to increase the overall electro-energetic system resilience.

How to cite: Abbate, A. and Mancusi, L.: A fast geo-hazard assessment for electro-energetic network systems using a simplified geo-hydrological model, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-22521, https://doi.org/10.5194/egusphere-egu24-22521, 2024.

The framework proposed by Piciullo et al., 2022 for a Internet of Things (IoT)-based local landslide early warning system (Lo-LEWS) consists of four main components: monitoring, modelling, forecasting, and warning. It was applied to a steep natural slope in Norway, equipped with various hydrological and meteorological sensors since 2016. Volumetric water content (VWC) and pore-water pressure (PWP) sensors were installed in 2016 (Heyerdahl et al., 2018). A weather station was added in 2022 to measure climate variables: rainfall, relative humidity, wind speed, air temperature among others. The sensors and weather station regularly send data to NGIs IoT data platform (NGI Live), which stores and makes the data available real-time through online dashboards and Application Programming Interface (API). GeoStudio software was used to create a reliable digital twin of the slope with the aim of back-calculating the in-situ hydrological conditions. Calibration, climate variables, and vegetation proved crucial for accurately modelling the slope's response . Sensitivity analysis on hydraulic conductivity and permeability anisotropy improved input data and model fitting. The hydrological model adequately represented monitored conditions up to a 1-year period (Piciullo et al., 2022). 

A fully operational IoT-based slope stability analysis has been recently established. The digital twin model has been used to evaluate the slope stability (i.e., factor of safety, FS) coupling SEEP and Slope analyses for 5 different 1-year datasets. Both past and future scenarios have been considered:  2019-2020, 2021-2022, 2022-2023, 2064-2065, 2095-2096. The inputs (i.e., hydrological and weather variables) and the FS results have been used to train different machine learning and statistical models. The feature considered are VWC, PWP, rainfall, temperature, LAI; the target was the FS. The best models able to predict the FS, given the features, are polynomial regression and random forest.

In order to predict the FS for the upcoming three days, PASTAS model (Collenteur et al., 2019) and the Norwegian Meteorological Institute webpage have been used to respectively forecast the hydrological variables (i.e., VWC and PWP) and rainfall, air temperature and relative humidity data. We created a web service that once a day automatically (1) fetches measured data from NGI Live using the NGI Live API, (2) runs predictions for the next three days based on the measured data, (3) sends the predicted values back to NGI Live, making them available for real-time visualization in online dashboards. This case study can be seen as a fully operational example of the use of IoT and digital twinning to provide a real-time stability assessment for a slope as well as a collaborative effort among different expertise: geotechnical, hydrological, instrumental and informatics.  

REFERENCES

Heyerdahl H., et al. (2018). Slope instrumentation and unsaturated stability evaluation for steep natural slope close to railway line. In UNSAT 2018: The 7th International Conference on Unsaturated Soils.

Collenteur R. A., et al. (2019). Pastas: Open Source Software for the Analysis of Groundwater Time Series. Groundwater, 57(6):877–885. URL: https://doi.org/10.1111/gwat.12925, doi:10.1111/gwat.12925.

Piciullo, L., et al. (2022) A first step towards a IoT-based local early warning system for an unsaturated slope in Norway. Nat Hazards 114, 3377–3407 (2022). https://doi.org/10.1007/s11069-022-05524-3 

How to cite: Piciullo, L., Abraham, M. T., Drøsdal, I. N., Paulsen, E. S., Capobianco, V., and Heyerdahl, H.: A fully operational IoT-based slope stability analysis for an unsaturated slope in Norway, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7275, https://doi.org/10.5194/egusphere-egu24-7275, 2024.

Climate change impacts the frequency and intensity of extreme weather events, leading to an increase in natural disasters globally. Heavy rainfall is a notable extreme weather event, acting as an external factor for landslides. In South Korea, where approximately 70% of the terrain is mountainous, the susceptibility to landslides is high. Despite the development and implementation of landslide early warning systems by the Korea Forest Service for local governments, the extent of landslide damage has been significant, reaching approximately 2,345 hectares in the last five years. Especially, last year, landslides occurred more than 800 times with severe human costs. The current early warning system, which focuses on administrative boundaries, has limitations in accurately identifying high-vulnerability landslide areas. To address this issue, this study introduces a landslide diagnostic model designed to assess the daily susceptibility of South Korea with fine spatial resolution. The model employs a semi-automated process that encompasses the acquisition of short-term climate forecast data and the generation of daily landslide susceptibility maps. The core algorithm of the model is based on the random forest method, predicting susceptibility at a spatial resolution of 100 meters. The model integrates various feature datasets, including meteorological, topographic, and land surface data, which are closely linked to landslide occurrences. The training model utilized landslide inventory data from 2016 to 2022, with various performance indicators employed for calibration and validation. Additionally, the landslide inventory data from 2023 was utilized for final model verification. Notably, the model incorporates a 3-day climate forecast data process provided by the Korea Meteorological Administration, enabling the prediction of short-term daily landslide susceptibility. This landslide diagnostic model holds the potential to enhance landslide prevention and preparedness at both local and regional scales.

How to cite: Lee, S., Roh, M., Kim, U., and Lee, W.-K.: Machine learning-based landslide susceptibility mapping for short-term risk assessment in South Korea, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7357, https://doi.org/10.5194/egusphere-egu24-7357, 2024.

As a result of extreme weather conditions such as heavy precipitation, natural slopes can fail dramatically. While the pre-failure deformation is sometimes apparent in retrospect, it remains challenging to predict the sudden transition from gradual deformation to runaway acceleration. Recent advancements in remote sensing techniques, like satellite radar interferometry (InSAR), enable high spatial and temporal resolution measurements of deformation and topographic information, providing valuable insights into landslide detection and activity. 

Landslides are common on the Big Sur coast, Central California, USA due to active tectonics, mechanically weak rocks, and high seasonal precipitation. We use satellite InSAR data from Copernicus Sentinel-1A/B to identify 23 active landslides within our 175 km² study site; one is Mud Creek, a slow-moving, deep-seated landslide that catastrophically failed in May 2017 and another is Paul’s Slide, which has experienced nearly constant motion for decades. 

We use multilayer networks to investigate the spatiotemporal patterns of slow deformation on the 23 active landslides. In our analysis, we transform observations of the study site — ground surface displacement (InSAR) and topographic slope (digital elevation model) — into a spatially-embedded multilayer network in which each layer represents a sequential data acquisition period. We use community detection, which identifies strongly-correlated clusters of nodes, to identify patterns of instability. We have previously shown [Desai et al., Physical Review E, 2023] that using high-quality data containing information about the fluidity (via velocity as a proxy) and susceptibility (slope) of the area successfully forecasts the transition of the Mud Creek landslide — the only formally slow-moving landslide in this collection to have catastrophically collapsed — from stable to unstable. 

Using multivariate analysis, we compare the traits of the active landslides, such as precipitation, vegetation, deformation, topography, NDVI, and radar coherence, against the results of the community detection. A strong indicator of instability is a combination of poor InSAR coherence and high displacement. Combined with community detection, we are able to differentiate between creeping landslides that are stable and landslides that display concerning trends that may warn of catastrophic failure.

How to cite: Desai, V. D., Handwerger, A. L., and Daniels, K. E.: Evaluating Landslide Susceptibility on the Big Sur Coast, California, USA using Complex Network Theory, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14020, https://doi.org/10.5194/egusphere-egu24-14020, 2024.

Recent works on landslide displacement forecasting using machine learning or deep learning models show relevant performance. However, they are mostly based on the use of historical displacement information and do not provide information on the most predictive features in terms of meteorological and hydrogeological variables for the forecast, and thus the identification of possible precursory factors. In this context, providing approaches based on EXplainable Machine Learning (XML) is essential for landslide forecasting as it concerns making decisions about risk mitigation actions, it supports the identification of possible precursory factors and it increases confidence in the predictions.
The proposed XML-based landslide forecasting approach is developed and tested using ensemble learning methods such as Random Forest and XGBoost. It relies on the use of multi-year and multi-parameter data chronicles to analyse the relationships between surface displacements (target data) and hydro-meteorological conditions (predictor data). Displacement and meteorological data are acquired through the landslide monitoring network. Hydrological data, when not available, are simulated discharge calculated with reservoir based-model; the simulations allow to construct water level time series for each water reservoirs identified in the unstable slope.
The predictive time series are decomposed into a set of 340 descriptive features (mean, variance, difference, number of rainy days, number of consecutive rainy periods of X days, …). The displacement time series are detrended using the multiplicative decomposition method.
This method has been applied to several use cases, such as the Séchilienne landslide located southwest of the Belledonne massif (French Alps). The Random forest and XGBoost models are trained and tested over periods of 12 and 5 years respectively, and applied to three automatic extensometers located in the most active part of the landslide. The results indicate that the main features used include variations in water levels over past 10 to 30 days, as well as the number of consecutive rainy period during the month. These results are associated with accurate predictions for the three extensometers, with coefficients of determination ranging between 0.37 and 0.46.
We show that these models have high predictive power while informing about the most important hydro-meteorological features. The application of the models to trendless displacement time series significantly improves prediction accuracy.

How to cite: Maillard, O., Bertrand, C., and Malet, J.-P.: Forecasting landslide motion with EXplainable Machine Learning models: the use case of Séchilienne landslide (French Alps) to identify the relevant predicting variables, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16825, https://doi.org/10.5194/egusphere-egu24-16825, 2024.

The broad family of shallow flow models arises from depth-averaging the underlying governing balance laws. Depth-averaging yields an analytical model complexity reduction, increasing computational efficiency and reducing the number of model parameters. Consequently, shallow flow models become a desirable choice for various scientific and engineering applications, such as landslide prediction and coastal engineering. In the realm of landslide modelling, different variants of shallow flow models are often tailored - sometimes in an ad hoc manner - to specific physical phenomena, such as basal shear, non-hydrostatic effects, kinetics, or phase change processes. Therefore, selecting the most appropriate shallow flow model for a particular scenario based on quantitative reasoning poses a formidable challenge. Quantifying the uncertainty associated with this model selection is essential to assess the reliability of the predictions of these shallow flow models.

Here, we present a unified Bayesian model selection workflow leveraging Gaussian Process emulation — a machine learning technique used for non-intrusive physics-based machine learning. It starts with model calibration, where we generate posterior samples. These are then used to calculate the marginal likelihood, the basis for our model selection. This process faces two computational bottlenecks: significant computational costs involved in numerous model evaluations during calibration and high-dimensional, intractable integrals in the computation of Marginal Likelihood. To address the former, we integrated Gaussian process emulators into the workflow using PSimPy, our in-house Python package, for predictive and probabilistic simulations. For the latter bottleneck, we conducted a comprehensive literature review, with particular emphasis on marginal likelihood computation techniques based on Importance Sampling and implemented single proposal density schemes and integrated them into the workflow.

We demonstrate our approach using elementary landslide runout models across varying fidelity levels, investigating the impact of data representation—specifically, comparing point data to time series data—while considering data characteristics such as velocity and distance. Additionally, we calibrated the discrepancy parameter for robust handling of uncertainties associated with the data. Our future work will focus on implementing advanced importance sampling schemes to enhance the computation of the Marginal Likelihood, especially in high-dimensional scenarios. Furthermore, emphasis will be placed on adopting a hierarchical approach to address data uncertainty in conjunction with model inadequacy, which is not accounted for in the existing workflow.

How to cite: Kumar, V. M. and Kowalski, J.: A unified Bayesian model selection workflow for geophysical free-surface flow, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18847, https://doi.org/10.5194/egusphere-egu24-18847, 2024.

Road infrastructure plays a key role in the economic development of a society. Thus, ensuring its functionality and safety conditions over time is a crucial and, at the same time, a demanding task that central and local authorities are asked to address. In Italy, road networks often develop within complex geological contexts, where active slow-moving landslides may generate risks to traveling persons and to the roads themselves, the latter being associated with socio-economic impacts. The identification of the road sections most exposed to landslide risk is critical for reducing the population potentially exposed to risk and for minimizing the repair/replacement costs. However, studies specifically oriented to roads affected by existing slow-moving landslides are quite rare in the scientific literature. This is possibly due to different reasons: landslide inventories with reliable information on the past and current state of activity of the phenomena are often not available; assessing the temporal probability of landslides characterized by a given intensity over large areas is not straightforward; the development of large datasets of road displacements and damage through traditional techniques can be time-consuming and sometimes not affordable.

This study proposes a conceptual model aimed at classifying the level of exposure to slow-moving landslide risk of stretches of roads at municipal scale. The activities have been developed in the context of the “Mitigation of natural risks to ensure safety and mobility in mountain areas of Southern Italy” (MitiGO) project.  Adopting a matrix-based approach, the following data are combined: landslide inventories, thematic information, displacement measurements derived from the interferometric processing of synthetic aperture radar images (DInSAR) and damage records obtained from Google Street View. First, a statistical model based on the bivariate correlations between the independent variables (i.e., each significant spatial variable derived from the thematic maps) and the dependent variable (i.e., the slow-moving landslides inventoried in the official map) is applied for zoning the susceptibility to slow-moving landslides at the municipal scale. Then, the information is combined with the level of damage and a monitored rate of movement based on DInSAR-derived ground-displacement measurements along the road network. The output is a correlation matrix combining all the information and classifying each stretch of the road network.

The proposed procedure has been applied to different access routes from a major regional road, the SS407 Basentana highway, to some urban centers of municipalities located in the Basento river basin (Basilicata region, southern Italy).

The analyses carried out at a municipal scale allow the classification of the road stretches potentially exposed to slow-moving landslide risk adopting a fairly simple qualitative ranking procedure, reliable in relation to the scale of analysis, which is based on a few data that are relatively easy to retrieve and to manage. The obtained results can be used to support studies of road networks over large areas aimed at the prioritization of risk-mitigation measures, as well as at the identification of road sections requiring further geomorphological surveys and geotechnical analyses, to be conducted in more detail at a larger scale.

How to cite: Pecoraro, G., Nicodemo, G., Menichini, R., Luongo, D., Peduto, D., and Calvello, M.: Analyses of slow-moving landslides interacting with the road network: case studies in Basilicata region (southern Italy), EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10593, https://doi.org/10.5194/egusphere-egu24-10593, 2024.

How to cite: Smith, H., Neverman, A., Betts, H., and Spiekermann, R.: The influence of rainfall patterns on shallow landslides in New Zealand, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1696, https://doi.org/10.5194/egusphere-egu24-1696, 2024.

The common understanding of landslides points to intense precipitation as a primary trigger. However, this explanation falters when considering landslides occurring with minimal or no rainfall, challenging the basis of empirical and numerical analyses. Taking advantage of a dataset documenting 1,118 landslide disasters with casualties in China since 1984, this study incorporates field investigations, laboratory experiments, and numerical simulations to unravel the mechanisms behind the delayed initiation of landslides influenced by post-precipitation runoff and infiltration. A noteworthy finding emerges: over 75% of catastrophic landslides in China exhibit a temporal delay compared to triggering rainfalls, typically manifesting within one week following peak precipitation. The temporal dynamics of precipitation-induced landslide delays show a range from months to hours, with the delay positively correlated to both landslide scale and the severity of regional drought. Spatially, delayed landslides are frequently related to runoff recharge by upstream catchment, playing a pivotal role in the initiation process. Consideration of topography, climate, and human activities leads to the identification of four typical runoff recharge patterns. We use such patterns to investigate the relationships with the upstream catchment area and delay time, influenced by surface runoff migration and supplied runoff infiltration. Hydrological and slope stability calculations underscore the significance of the catchment area to landslide area ratio while delay time is predominantly governed by surface runoff migration and supplied runoff infiltration into the sliding soil. Results unveil a consistent sequence: robust runoff recharge facilitates water infiltration into weak rock fractures or soil mass, resulting in a gradual increase of pore water pressure. This sequence culminates in the delay of landslide initiation compared to the peak precipitation. These findings may contribute to a scientific foundation for early warning and prediction related to such landslides, thereby mitigating associated risks.

How to cite: Wu, K., Simoni, A., and Chen, N.: Understanding Delayed Landslides: A Study of 1,118 Fatal Incidents in China Influenced by Post-Precipitation Runoff, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2680, https://doi.org/10.5194/egusphere-egu24-2680, 2024.

Deep Learning (DL) algorithms have demonstrated superior efficacy compared to traditional Machine Learning (ML) methods in the realm of landslide detection through the analysis of Remote Sensing (RS) imagery. However, their performance is notably contingent upon the quantity of manual annotations utilized during the training process. This investigation delves into the utilization of two distinct Self-Supervised Learning (SSL) models, specifically the Simple Framework for Contrastive Learning of Visual Representations (SimCLR) and Swapping Assignments between multiple Views (SwAV). These models were adapted and enhanced for downstream tasks, particularly in the domain of landslide detection. To train the SSL models, the Landslide4Sense competition dataset was employed, consisting of 3799 training patches, 245 validation patches, and 800 testing patches generated from Sentinel-2 images acquired from diverse regions worldwide. During the training of SimCLR and SwAV models, only the training patches were utilized, with a series of data augmentations applied to the input dataset based on each model's architecture. Both models employed ResNet-50 as the encoder.

For the downstream task of landslide detection, a custom U-Net model was developed. The trained ResNet-50 served as the encoder, and during fine-tuning, only the decoder part was permitted to be trained while the encoder remained frozen. During the fine-tuning process, subsets comprising 1% and 10% of labeled data from the training dataset were randomly selected to train the model, and predictions were exclusively conducted on the testing data. While a conventional supervised ResU-Net model, which was trained on all labeled training datasets, attained an F1 score of 72%, the SSL models achieved F1 scores of 64% and 71% with 1% labeled data, and 68% and 76% with 10% labeled data for SimCLR and SwAV, respectively. In addition, comparisons were conducted with all supervised reference models in the Landslide4Sense competition, revealing that SwAV, with 10% labeled data, outperformed all models, surpassing their top model by 4%. This study underscores the potential of SSL techniques in the segmentation and classification of RS images for natural hazard mapping, particularly in scenarios where labeled data is not available or is limited.

How to cite: Shahabi, H., Ghorbanzadeh, O., Homayouni, S., and Ghamisi, P.: A Comparison of SimCLR and SwAV Contrastive Self-Supervised Learning Models For Landslide Detection, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4772, https://doi.org/10.5194/egusphere-egu24-4772, 2024.

Strong earthquakes on mountain slopes can trigger numerous landslides, a significant secondary hazard responsible for a substantial proportion of fatalities in the affected area. In this study, we present a model framework for rapidly creating coseismic landslide probability distribution maps using machine learning models and optimal conditioning factors. To illustrate our approach, we focus on the case of the Mw 7.2 Haiti earthquake in 2021 and predict the distribution of coseismic landslides based on historical landslide data collected following the Mw 7.0 Haiti earthquake in 2010. To validate our findings, we mapped all the landslides triggered during the 2021 event. Furthermore, we conduct a comparative analysis of various landslide-conditioning factors (seismic, topographic, lithologic, and hydrological variables) in relation to the coseismic landslides occurring during both earthquake events in 2010 and 2021, to reassess the factors feed into the machine learning model. We observed noticeable differences in patterns of several conditioning factors between the two events EQIL distributions (e.g., tectonic and releif factors), but consistent similarities in other terrain factors (e.g., slope, curvature, topographic wetness index, etc.). Our Random Forest (RF) model, initially trained using the 2010 landslide inventory and 15 selected factors, effectively predicts 2021 landslides with an area under curve (AUC) score of 0.83. Improved performance is achieved when we use a reevaluated set of six factors for training, resulting in an AUC score of 0.90, with  93% of landslides falling into the high to medium probability class. These findings demonstrate the feasibility of rapidly generating highly accurate coseismic landslide distribution maps, even when there are considerable differences in key conditioning factors, highlighting the applicability of ML models to complex problems.

How to cite: Thanveer, J. and Pulpadan, Y. A.: Rapid Estimation of Earthquake Induced Landslides using Machine Learning Models: Insights from Haiti Earthquakes, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-16454, https://doi.org/10.5194/egusphere-egu24-16454, 2024.

The wide physiographic variability and the abundance of rainfall and landslide data make Italy an appropriate site to study variations in the rainfall conditions responsible for triggering landslides.

For more than two decades, the Research Institute for Geo-Hydrological Protection of the Italian National Research Council (CNR-IRPI) has been carrying out a specific research activity aimed at collecting information on rainfall-induced landslides in Italy. The information comes mainly from chronicle sources (newspapers in print or electronic format, websites, etc.) and institutional sources (reports on interventions carried out by the Fire Brigade and other institutional entities following reports of weather-induced landslides). The information collected has been used to compile the ITAlian rainfall-induced LandslIdes CAtalogue (ITALICA), freely accessible at https://zenodo.org/records/8009366. A description of the main features of the catalogue and the procedures adopted to fill it out can be found at https://essd.copernicus.org/articles/15/2863/2023/.

ITALICA, which is being continuously updated, to date contains data on more than 6300 rainfall-induced landslides that occurred in Italy during the period 1996-2021. The peculiarity and specificity of the catalogue lies in the mastery and control of the landslide records, which have very high levels of spatial and temporal accuracy. In particular, for more than one third of the catalogue, landslides are spatially and temporally localized with an uncertainty of less than one km² and one hour, respectively. The availability of accurate and up-to-date information on the geographic location and time of onset of landslides is essential for improving the predictive ability of landslides. Different subsets of the catalogue have been already used to calculate national and regional rainfall thresholds implemented in early warning systems in Italy.

The first published version of ITALICA did not contain information on the rainfall conditions associated with the landslides. In the new release, presented here, we add the cumulate rainfall, rainfall duration and mean rainfall intensity values of the rainfall conditions responsible for the failures listed in the catalogue. The rainfall conditions are reconstructed by means of the CTRL-T automatic tool (https://zenodo.org/records/4533719) and using hourly rainfall measurements from more than 3000 rain gauges distributed over the Italian territory. Rainfall records are provided by the Italian National Department for Civil Protection. The spatial and temporal features of the reconstructed landslide-triggering rainfall conditions are analysed in depth.

Given the rising demand for high-quality data to be used in comprehensive analyses and data-driven models, this dataset might be very useful for assessing the rainfall triggering conditions of landslides in Italy, either by empirical or physically based models. In particular, we expect our results to have an impact on the definition of new rainfall thresholds to be implemented in landslide early warning systems at regional and national scales.

Work financially supported by the Italian National Department for Civil Protection (Accordo di Collaborazione 2022-2024) and the PRIN-ITALERT project (PRIN2022 call, grant number: 202248MN7N, funded by NextGenerationEU).

How to cite: Peruccacci, S., Gariano, S. L., Melillo, M., Guzzetti, F., and Brunetti, M. T.: Determination and analysis of the rainfall triggering landslides in the ITALICA catalogue , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7330, https://doi.org/10.5194/egusphere-egu24-7330, 2024.

Understanding landslide failure processes is pertinent to predict and minimize the effects of landslides. A variety of elements, such as geology, topography, and soil conditions, can lead to slope failures triggered via natural causes e.g., rainfall and earthquakes, setting off the failure movements. Proper geotechnical analysis requires knowledge of both the triggering event and the subsequent movement patterns of the landslide. This information is vital for accurately predicting when and where landslides might occur. To integrate this information into existing landslide inventories, we introduce Landsifier 2.0, a tool designed to meet the needs of the landslide research community. This Python-based library allows seamless usage of machine learning models to extract information regarding landslide triggers and failure movements solely based on inventories of landslides. Powered by topology, a high-dimensional feature extraction module encapsulated within our library, information accessed via a landslide's shapes and configurations allows the identification of triggers (e.g., earthquake-and rainfall-triggered landslides) and failure movements (e.g., rotational slides, translational slides, debris flows, rock falls) of undocumented landslide inventories through continuous remote sensing missions. We showcase the library’s application in diverse geomorphological and climatic settings e.g., South-western China, Denmark, Turkey, Japan, Italy and more. We anticipate that Landsifier 2.0 will be particularly useful in the predictive modelling domain (including susceptibility and hazard modelling) of landslide studies, where precise information about triggers and failure dynamics is essential for developing reliable predictive models.

References:
Rana, Kamal, Uğur Öztürk, and Nishant Malik. 2021. “Landslide Geometry Reveals Its Trigger.” Geophysical Research Letters 48(4). doi: 10.1029/2020gl090848.
Rana, Kamal, Nishant Malik, and Uğur Öztürk. 2022. “Landsifier v1.0: A Python Library to Estimate Likely Triggers of Mapped Landslides.” Natural Hazards and Earth System Sciences 22(11):3751–64. doi: 10.5194/nhess-22-3751-2022.
Rana, Kamal, Kushanav Bhuyan, Joaquin Vicente Ferrer, Fabrice Cotton, Uğur Öztürk, Filippo Catani, and Nishant Malik. 2023. “Landslide Topology Uncovers Failure Movements.” arXiv (Cornell University). doi: 10.48550/arxiv.2310.09631.

How to cite: Nava, L., Bhuyan, K., Kapoor, M., Rana, K., Rosi, A., Vicente Ferrer, J., Ozturk, U., Floris, M., van Westen, C., and Catani, F.: Landsifier 2.0: Towards automating landslide trigger and failure movement identification, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15580, https://doi.org/10.5194/egusphere-egu24-15580, 2024.

Landslides are a serious geologic hazard common to many countries around the world.  They can result in fatalities and the destruction of infrastructure, buildings, roads, and electrical equipment. Especially rapid-moving landslides, which occur suddenly and travel at high speeds for miles, can pose a serious threat to life and property. Landslide inventories are essential to understand the evolution of landscapes, and to ascertain landslide susceptibility and hazard, and it can be of help for any further hazard and risk analysis. Although  many landslides inventories have already been created worldwide, often these archives of historical landslide events  lack precise information on the date of landslide occurrence. Many of these inventories also lack completeness especially in case of smaller landslides which is also caused by  landslides erosion processes, human impact, and vegetation  regrowth. Precise determination of landslide occurrence time is a big challenge in  landslide research. Optical and Synthetic Aperture Radar (SAR) images with multi-spectral and textural features, multi-temporal revisit rates, and large area coverage provide opportunities for landslide detection and mapping. Landslide-prone regions are frequently obscured by cloud cover, limiting the utility of optical imagery. The capacity of SAR sensors to penetrate clouds allows the use of SAR satellite data to provide a more precise temporal characterization of the occurrence of landslides on a regional scale. The archived Copernicus Sentinel-1 satellite, which has a 6 to 12-day revisit period and covers the majority of the world's landmass, allows for more precise identification of landslide failure timings. The time-series of SAR amplitude, interferometric coherence, and polarimetric features (alpha and entropy) have strong responses to landslide failures in vegetated regions. This is characterized by a sudden increase or decrease in their values. Consequently, the abrupt shifts in the time-series of SAR-derived parameters, triggered by the failure, can be recognized and regarded as the failure occurrence time. The aim of this study is to determine the time period of failure occurrences by automatically detecting abrupt changes in the time series of SAR-derived parameters. We present a strategy for anomaly detection in time-series based on deep-learning to identify the failure time using four parameters derived from SAR time series. In this strategy, we introduce a gated relative position bias to an unsupervised Transformer model to detect anomalies in a multivariate time-series composed of four SAR-derived parameters. We conduct an experiment involving multiple landslides and compare the performance of our proposed strategy for detection of the failure time period with that of the LSTM model. Our strategy successfully identifies the time of landslide failure, which closely approximates the actual time of occurrence when compared to the LSTM model employed in this study.

How to cite: Wang, W., Motagh, M., Xia, Z., Plank, S., Li, Z., Orynbaikyzy, A., Zhou, C., and Roessner, S.: A deep Learning-based approach for landslide dating from time-series of SAR data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12097, https://doi.org/10.5194/egusphere-egu24-12097, 2024.