Detecting volcanic deformation in Hawai‘i using trustworthy multimodal deep learning techniques

Abstract Monitoring volcanoes involves a variety of data sources and methods to maintain complete continuity of coverage. Global navigation satellite system (GNSS) and interferometric synthetic aperture radar (InSAR) are commonly used complementary methods to assess the deformation state of a volcano as magma migrates beneath the surface. The amount of data these methods produce, however, is growing rapidly beyond human analysis capabilities and is becoming difficult to manage. Here, we create a novel multimodal deep learning framework to ingest InSAR and GNSS data simultaneously and classify the deformation state of the system. We apply this methodology to Mauna Loa, Hawai‘i given its wealth of InSAR and GNSS data as well as its propensity to deform on multiple timescales. Our model performs with high accuracy and is able to identify both slow and fast deformation from 2015 to 2023. The multimodal nature of our model also allows us to identify the presence of atmospheric noise in InSAR data. Furthermore, we employ explainability algorithms to show that our model is making decisions for the right reasons and to connect complex black-box machine learning mappings to current real-world geodetic interpretations of the Mauna Loa magmatic system.