EGU25-17662, updated on 15 Mar 2025
https://doi.org/10.5194/egusphere-egu25-17662
EGU General Assembly 2025
© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.
Automated Site Effects Mapping in Mayotte Using Airborne Electromagnetic Data and Machine Learning
Cécile Gracianne, Hugo Breuillard, Célia Mato, Pierre-Alexandre Reninger, Agathe Roullé, Anne Raingeard, and Roxanne Rusch
Cécile Gracianne et al.
  • BRGM, Risks, Orléans Cedex 2, France (c.gracianne@brgm.fr)

Recent seismic hazard assessments in Mayotte have highlighted the island's significant exposure to site effects during earthquakes. These effects are closely linked to its complex geological setting, characterized by altered volcanic formations whose heterogeneous geometry leads to strong spatial variations in ground motion. In response to governmental requests, a site effects map is being developed to raise public awareness and support risk-informed urban planning.

A novel methodology for site effects mapping has recently been developed at BRGM, integrating airborne electromagnetic (AEM) data with borehole logs, geological maps, and seismic data (MASW and H/V measurements). This approach was tested on three test sites covering 12 km² of Mayotte surface, and it has demonstrated its potential in imaging the geological interfaces responsible for site effects. However, the current methodology relies on expert-driven data interpretation, making its large-scale application highly labour-intensive and costly. To overcome this limitation, partial automation of the data processing is required in order to handle larger datasets efficiently.

Machine learning techniques offer a promising solution to address this challenge. The test sites provided a unique training dataset, associating resistivity profiles derived from AEM data with the position of geological interfaces responsible for site effects within the soil column. These interface locations were determined through the integration and interpretation of all available geological and geophysical data, including MASW, H/V measurements, and borehole logs. Using this dataset, we trained various models, including Random Forest and Convolutional Neural Networks (CNN), to predict the localization of geological interfaces responsible for site effects based on AEM data.

Preliminary results indicate that the CNN model shows good performances on this task. Nevertheless, further improvements require the expansion of training datasets, underscoring the significant investment needed to generalize this approach to other regions. Future research will focus on refining predictive models and optimizing data acquisition to support large-scale implementation.

How to cite: Gracianne, C., Breuillard, H., Mato, C., Reninger, P.-A., Roullé, A., Raingeard, A., and Rusch, R.: Automated Site Effects Mapping in Mayotte Using Airborne Electromagnetic Data and Machine Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17662, https://doi.org/10.5194/egusphere-egu25-17662, 2025.