Skip to content

Detecting Volcanic Deformation in Hawaiʻi Using Multimodal Deep Learning Techniques

Tyler Paladino1, Emily Montgomery-Brown1, Michael Poland1,2, Marco Bagnardi1 , Lopaka Lee3

  • Affiliations: 1Cascades Volcano Observatory, U.S. Geological Survey, Vancouver, WA, USA;[2]{.mark}Yellowstone Volcano Observatory, U.S. Geological Survey, Vancouver, WA, USA; 3Advanced Research Computing, U.S. Geological Survey, Sioux Falls, SD, USA

  • Presentation type: Poster

  • Presentation time: Monday 16:30 - 18:30, Room Poster Hall

  • Poster Board Number: 132

  • Programme No: 3.1.24

  • Theme 3 > Session 1

Abstract

InSAR and GNSS are important for volcano monitoring and characterizing subsurface magmatic plumbing systems. Dense GNSS networks exist at many volcanoes throughout the world, and InSAR data are becoming more ubiquitous with each new SAR satellite launched. Machine learning has previously been used to automatically detect when and where deformation is occurring in large volumes of InSAR data. Here, we build upon these techniques by showcasing a multimodal deep learning model that utilizes both InSAR imagery and GNSS timeseries to assess a volcano's deformation state. The model has two independent branches that process each datatype separately: the InSAR branch is a pretrained convolutional neural network (CNN) that is modified to process InSAR data; the GNSS branch is a simpler multi-layer perceptron (MLP) that is trained from scratch. These streams are then combined by averaging their output probabilities. In this manner, each branch can also be used independently. Also, while accuracy is important, we strive to make the reasoning behind the network's decision-making transparent to avoid the "black box" nature of many machine learning models by using integrated gradients backpropagation techniques. We apply this model to Mauna Loa, Hawaiʻi using Sentinel-1 InSAR data and GNSS data from the USGS Hawaii Volcano Observatory, both spanning 2015 to 2024. We chose Mauna Loa as it deforms often and is a very high threat volcano. Our aim is to develop a trustworthy multimodal model that can automatically detect deformation using a diversity of spatiotemporal datasets to assist in monitoring and science efforts.