Developing Machine Learning for Geoscientific Data: Understanding and Monitoring Natural Hazards

Ponente: Dra. Sophie Giffard-Roisin, Université Grenoble Alpes / ISTerre, IRD, Francia

Resumen:
Predicting and monitoring natural hazards requires processing massive, noisy, and multimodal spatiotemporal data streams. This talk explores the development of tailored AI architectures designed to overcome key mathematical and algorithmic challenges in geosciences, particularly the severe lack of ground-truth labels.

The presentation will focus on three main machine learning paradigms:
 

• Supervised Learning & XAI: Utilizing Random Forests and Convolutional Neural Networks (CNNs)—coupled with Explainable AI techniques—for rapid earthquake damage estimation, tropical storm forecasting, and seismic waveform analysis.
• Physics-Driven Synthetic Training: Overcoming label scarcity by training deep neural networks on realistically simulated data. Applications include active fault characterization using CNNs, earthquake ground deformation estimation from satellite imaging with physically-regularized U-Nets (GeoFlowNet), and slow slip event detection in GPS time series via Spatiotemporal Graph Neural Networks.
• AI Foundation Models: Leveraging the recent shift toward massive, pre-trained architectures (e.g., Vision Transformers) and efficiently fine-tuning them for complex, low-data visual tasks, demonstrated through the semantic segmentation of volcanic plumes.

Transmisión en vivo