Deep learning detection and quantification of volcanic thermal signals in infrared satellite data

Abstract

Volcanic eruptions pose hazards to human lives and livelihoods (Loughlin et al., 2015). To mitigate these hazards, volcano monitoring groups aim to detect signs of unrest and eruption as early as possible. Prior to eruption volcanoes may show various signals of unrest, including: increased surface temperatures, surface deformation, increased seismicity, increased degassing, and more. Here we focus on one approach to monitor volcanic unrest: detecting high-temperature localized volcanic heat emissions, otherwise known as hotspots. The presence of hotspots can indicate subsurface and surface volcanic processes that precede, or coincide with, eruptions. Space-borne infrared sensors can identify hotspots in near-real-time; however, automatic hotspot detection systems are needed to efficiently analyze the large quantities of data produced. While hotspots have been automatically detected for over 20 years with simple thresholding algorithms, new computer vision technologies, such as convolutional neural networks (CNNs), enable improved detection capabilities. Here we introduce HotLINK: the Hotspot Learning and Identification Network, a CNN-based model to detect volcanic hotspots in VIIRS (Visible Infrared Imaging Radiometer Suite) imagery. We find that HotLINK achieves an accuracy of 96% when evaluated on a validation dataset of ~1,700 unseen images from Mount Veniaminof and Mount Cleveland volcanoes, Alaska, and 95% when evaluated on a test dataset of ~3,000 images from six additional Alaska volcanoes (Augustine Volcano, Bogoslof Island, Okmok Caldera, Pavlof Volcano, Redoubt Volcano, Shishaldin Volcano). Additional testing on ~700 labeled MODIS images demonstrates that our model is applicable to this sensor's data as well, achieving an accuracy of 98%. We apply HotLINK to 10 years of VIIRS data and 22 years of MODIS data for the eight aforementioned Alaska volcanoes. From these time series we find that HotLINK accurately characterizes background and eruptive periods, similar to a threshold-based method, MIROVA, but also detects more subtle warming signals, potentially related to volcanic unrest. In particular, analysis of the Mount Veniaminof record demonstrates that HotLINK is able to detect subtle hotspot signals that are coincident with elevated seismicity, potentially indicative of surface heating due to shallow magma intrusion and/or degassing. We identify three advantages to our model over its predecessors: (1) the ability to detect more subtle volcanic hotspots and produce fewer false positives, especially in daytime imagery; (2) the incorporation of probabilistic predictions for each detection that provide a measure of detection confidence; and (3) its transferability to multiple sensors and multiple volcanoes without the need for threshold tuning, suggesting the potential for global application. HotLINK is able to identify eruptions and potentially precursory warming signals in infrared satellite data, making it a valuable tool for monitoring volcanoes and tracking their heat released over time.