Landslides and Climate Change in highly Sensible Environments: Seismology, Earth Observation and Artificial Intelligence – HighLand

HIGHLAND relied on an integrated processing chain combining seismology, remote sensing, and artificial intelligence, deployed on high-performance computing infrastructures. Its objective was to overcome the limitations of landslide detection by exploiting the complementarities between seismic signals and satellite observations.

In seismology, traditional approaches focused on long-period waves and could only detect the largest events. To cover the full spectrum, HIGHLAND developed methods exploiting high-frequency signals, generated by all landslides but heavily contaminated by noise. Machine learning played a central role in addressing this challenge. Supervised algorithms, such as random forests, were used to efficiently discriminate between landslides and other seismic sources. A continuous classification approach using sliding signal windows was implemented, bypassing the conventional preliminary detection step, which often leads to missed events.

To improve generalization and the detection of atypical events, HIGHLAND explored semi-supervised and unsupervised strategies capable of handling the high variability of sources and class imbalance. Particular attention was given to Self-Supervised Learning (SSL), which enables the direct use of large volumes of unlabeled data to learn discriminative signal representations. These methods proved capable not only of enhancing the detection of weak signals but also of identifying rare events, such as the Dickson Fjord landslide and the subsequent tsunami, over eight years of continuous data. This methodological advance represented a decisive step toward comprehensive exploration of seismic archives.

All these methods required processing massive volumes of data from thousands of seismic stations. Their implementation on the high-performance computing center at the University of Strasbourg significantly reduced computation times and made near real-time applications feasible.

The HIGHLAND project led to several major advances in the study and monitoring of seismogenic landslides.

The first outcome was the development of a processing chain based on Self-Supervised Learning (SSL) combined with agglomerative clustering. This approach enabled, for the first time, the exhaustive exploration of decades of seismic records and the automatic identification of rare or atypical sources, such as the Dickson Fjord landslide and the associated tsunami. Conducted notably as part of J. Rimpot’s doctoral work, these studies represent a decisive methodological advance.

The second outcome was the implementation of a processing chain integrating automatic event localization and classification at the regional scale. Applied to 23 years of seismic data in the Alps, it led to the identification of nearly 1,000 new seismogenic landslides. Analysis of their temporal distribution revealed contrasting seasonal patterns depending on the mountain ranges, linked to the interaction between temperature, precipitation, snow cover, and snowmelt. A comparison before and after 2010 highlighted a rise in temperatures (0.5–1.1 °C), more pronounced precipitation seasonality, and a shift of gravitational activity toward earlier periods, in connection with climate change. These results were explored in depth, notably in C. Groult’s doctoral work.

The third outcome concerns the direct estimation of landslide dynamic properties from their seismic signals. Gradient boosting models were trained to predict volumes and run-out distances. The results show a median error of 42% for volumes (R² = 0.84) and 19% for distances, providing an alternative to classical inversion approaches and opening prospects for rapid hazard assessment in near real time.

Overall, HIGHLAND demonstrated the feasibility of exhaustive exploration of seismic archives and the relevance of artificial intelligence for building comprehensive and homogeneous catalogs. The project highlighted the role of climatic conditions in Alpine gravitational activity and proposed innovative methods for characterizing event dynamics. These results provide a solid foundation for the future development of automated landslide detection and monitoring systems.

The advances achieved by HIGHLAND open several major prospects for research and monitoring of gravitational instabilities.

The first concerns the implementation of near real-time automatic landslide detection. The processing chain developed and tested on several decades of data now needs to be optimized for continuous operation. The goal is to deploy this system across the Alps and then progressively extend it globally. Such a capability would allow monitoring of gravitational activity with the same responsiveness as for earthquakes and would constitute a valuable tool for risk anticipation.

A second prospect involves the development of large-scale self-supervised approaches. Their generalization should enable the exploration of extensive seismic archives in highly diverse contexts: the Alps, Himalayas, Arctic regions, volcanoes, or submarine environments. The ambition is to produce comprehensive and homogeneous catalogs covering a wide variety of sources and to better understand their relationships with climatic and hydrological forcings.

The creation of such catalogs paves the way for training foundation models for seismology, inspired by those already transforming language processing and image analysis. These models would be capable of learning robust and transferable representations of seismic signals, usable for multiple tasks: detection, classification, source discrimination, and dynamic property estimation. They would provide a unified framework to efficiently exploit the billions of hours of data available in global seismic databases. Such an approach would go beyond environmental seismology and could also benefit tectonic and volcanic seismology.

Finally, these developments fit within an international cooperation framework. Scaling up globally requires strengthening partnerships with seismic and satellite networks, as well as with operational agencies responsible for risk prevention. The expected benefits extend far beyond fundamental research: they directly concern mountain and remote-region risk management, climate change adaptation, and the protection of populations and infrastructure.

HIGHLAND has demonstrated the potential of artificial intelligence and high-performance computing to transform environmental seismology. The next steps consist in consolidating these achievements toward operational, global, and integrated monitoring systems, based on massive catalogs, self-supervised approaches, and foundation models, capable of addressing major scientific and societal challenges related to gravitational hazards in a warming climate.

The HighLand project proposes to combine seismology, remote sensing and machine learning to quantify the impact of climate on mass-wasting activity in regions of high latitude or altitude. The first objective of the project is the development of new processing chains to build, from the continuous recordings produced by regional seismological networks, instrumental catalogs of landslides. The systematic exploration of these seismological chronicles will be made possible by the use of machine learning algorithms and will enable the production of catalogs offering unparalleled spatio-temporal resolution. The seismological detection will be confronted with satellite observations with high temporal repetition possible thanks to the constellations of Sentinel and Landsat satellites. Three regions of the world will first be targeted by this new processing chain: Alaska, the Alps and Nepal. This multi-disciplinary approach will make it possible to produce the necessary observations and to build and constrain models to better understand the long and short-term links between climate and mass wasting activity. The prototype of the processing chain will serve as the basis for a system for observing and listening to the landside activity in near real time in these regions of the world and then on a global scale.