Unsupervised Anomaly Detection for Volcanic Deformation in InSAR Imagery

Abstract

Satellite-based Interferometric Synthetic Aperture Radar (InSAR) images have the potential to detect volcanic deformation prior to eruptions, but while a vast number of images are routinely acquired, only a small percentage contain volcanic deformation events. Manual inspection could miss these anomalies, and an automatic system modeled with supervised learning requires suitably labeled data sets. To tackle these issues, this paper explores the use of unsupervised deep learning on InSAR images for the purpose of identifying volcanic deformation as anomalies. We test three different state-of-the-art architectures, one convolutional neural network Patch Distribution Modeling (PaDiM) and two generative models (GANomaly and Denoising diffusion probabilistic models (DDPM)). We propose a preprocessing approach to deal with noisy and incomplete data points. We further improve the performance of PaDiM by using a weighted distance, assigning greater importance to features from deeper layers. The final framework was tested with five different volcanoes, which have different characteristics and its performance was compared against an existing supervised learning method for volcanic deformation detection. The experiments show that our final anomaly detection outperforms the supervised learning method, particularly where the characteristics of deformation are unknown. Our framework can thus be used to identify deformation at volcanoes without needing prior knowledge about the deformation patterns present there.