Antoine BRALET

Subject: Multimodal Deep Learning and Analysis of Spatio-Temporal Dynamics from Remote Sensing Images

Abstract: The ultimate goal of the research is to detect landslides on satellite images using deep learning methods. This study distinguishes between sudden ground movements caused by heavy rainfall and slow movements that are destructive over the long term. Several methods have been implemented: translation from radar to optical modality for cloud-independent detection, task combination to make translation more relevant, multimodal detection with missing modalities, creation of InSAR datasets, and explainability of neural networks are all topics that shape the thesis.

The topic of the research targets the detection of sliding areas from remote sensing images using deep learning approaches. Both sudden and slow-moving phenomena are of interest in the thesis, requiring the implementation of several deep learning techniques. Among them, the major contributions lie in radar-optical modality translation for weather-robust detection, leveraging land-cover classification to increase translation reliability, applying multimodal slide detection algorithms within a missing modality context, creating a new InSAR dataset, and introducing explainability within the networks.