Towards a generic clustering approach for building seismic catalogues from dense sensor networks

In the context of climate change, the occurrence of geohazards such as landslides or rockfalls might increase. Therefore, it is important to have the ability to characterise their (spatial and temporal) occurrences in order to implement protection measures for the potential impacted populations and infrastructures. Nowadays, several methods including Machine Learning algorithms are used to study landslides-triggered micro-seismicity and the associated seismic sources (eg. rockfalls and  slopequakes). Those innovative algorithms allow the automation of the processing chains used to build micro-seismicity catalogues, leading to the understanding of the landslide deformation pattern and internal structure. Unfortunately, each landslide context has its own seismic signature which requires the use of the most complete and handmade training samples to train a Machine Learning algorithm. This is highly time consuming because it involves an expert that needs to manually check every seismic signal recorded by the seismic network, which can be thousands per day.

The aim of this study is to develop semi-supervised and unsupervised clustering methods to characterise the micro-seismicity of landslides in near real time. Here, we present the preliminary results obtained for creating a landslide micro-seismicity catalogue from the analysis of a dense network of 50 seismic stations deployed temporarily at the Super-Sauze landslide (French Alps). First, we present the performance of supervised Random Forest and XGBoost trained models on the event signals. Then, an approach aimed at processing streams of raw seismic data based on 18s-length windows is explored. Finally, we discuss the clustering results and the transferability possibilities of the approach to other landslides and even environments (glaciers, volcanoes).