Modern Deep Learning Techniques for Volcanic Unrest Monitoring using InSAR Data

Satellites provide essential capabilities for widespread, regional, or global volcano surveillance, often offering the first indications of volcanic unrest or eruptions. Here, we focus on Interferometric Synthetic Aperture Radar (InSAR), a technology detecting surface deformation that is statistically strongly linked to volcanic activity. Recent technological advancements have enabled the generation of vast amounts of monitoring data—e.g., LiSC system currently provides over 3.4 million raw interferograms. Clearly, manual analysis of such a large dataset is no longer feasible. This talk presents several modern, learning-based techniques for ground deformation monitoring using InSAR data, including supervised, semi-supervised, and unsupervised learning approaches.

Supervised learning methods have successfully detected fringes in wrapped interferograms. We improved our CNN-based detection process [1,2,3] by incorporating state-of-the-art Transformers. However, these methods may miss ground deformations with characteristics differing from the training data. To address this limitation, we explore the potential of using semi-supervised learning [4]. In this approach, a global feature representation of InSAR data is learned through unsupervised contrastive learning [5], and the detection task is subsequently fine-tuned on a limited number of labelled samples. For unsupervised learning, our model identifies samples that deviate from the norm of the data as anomaly detection. It is performed in the feature space of unwrapped interferograms [6] and employs a statistical-based approach, Patch Distribution Modelling [7]. The results show that this method outperforms existing supervised learning techniques when the characteristics of deformation are unknown.

Interferograms capture deformation signals and atmospheric effects, which can distort detection accuracy. While GACOS provides atmospheric corrections, it may fail to fully remove effects and sometimes introduces artifacts. To address these limitations, we enhance our system with learning-based denoising techniques to mitigate atmospheric effects. Two approaches are presented: Transformer-based and diffusion model-based denoising. The first method adapts the state-of-the-art image denoising model, Reformer [8], but replaces the feed-forward network with multi-layer perceptron. The second method leverages Denoising Diffusion Probabilistic Models [9], incorporating turbulence noise in the forward diffusion process. Initial results, evaluated against GPS data, demonstrate that this method outperforms traditional time-series processing in mitigating atmospheric effects.

References:

[1] N Anantrasirichai et al., Application of Machine Learning to Classification of Volcanic Deformation in Routinely Generated InSAR Data. JGR Solid Earth, 2018

[2] N Anantrasirichai et al., A deep learning approach to detecting volcano deformation from satellite imagery using synthetic datasets, RSE, 2019

[3] N Anantrasirichai et al., The application of convolutional neural networks to detect slow, sustained deformation in InSAR time series, GRL, 2019

[4] N Anantrasirichai et al., Semi-supervised Learning Approach for Ground Deformation Detection in InSAR, Fringe, 2023

[5] T Yang et al., A Semi-supervised Learning Approach for B-line Detection in Lung Ultrasound Images. ISBI, 2023

[6] R Popescu et al., Anomaly detection for the identification of volcanic unrest in satellite imagery, ICIP, 2024

[7] T Defard et al., A Patch Distribution Modeling Framework for Anomaly Detection and Localization, ICPRW, 2021

[8] N Kitaev et al., Reformer: The Efficient Transformer, ICLR, 2020

[9] J Ho et al., Denoising diffusion probabilistic models. NIPS, 2020