Abstract: Interferometric Synthetic Aperture Radar (InSAR) is a powerful satellite-based tool for monitoring ground deformation, a key indicator of volcanic unrest. With modern systems generating millions of interferograms, manual inspection is infeasible. This talk introduces learning-based approaches—supervised, semi-supervised, and unsupervised—for scalable InSAR analysis. We extend traditional CNN-based fringe detection using Transformer architectures to improve accuracy. To enhance generalisation to unseen deformation types, we adopt a semi-supervised strategy that combines unsupervised contrastive learning with minimal labelled data. Additionally, our unsupervised method employs Patch Distribution Modelling to detect anomalies in unwrapped interferograms, outperforming supervised models when ground truth is limited. Moreover, as atmospheric noise remains a challenge, we address this using two learning-based denoising techniques: a modified Transformer (Reformer) and a turbulence-aware diffusion model. GPS-based evaluation shows our models outperform traditional methods, enabling more reliable, automated volcanic monitoring.
Bio: Dr Pui Anantrasirichai is an Associate Professor in Visual Computing at Computer Science, University of Bristol. She is actively involved in a range of projects focused on AI-based video production, image analysis and enhancement, medical imaging, and remote sensing. She is currently the Principal Investigator of two projects on video enhancement and underwater 3D reconstruction funded by EPSRC, and also leads a separate project on atmospheric turbulence mitigation funded by DASA.
