Automatic detection of changes in volcanic activity using ground based near-infrared cameras to monitor thermal incandescence

Abstract

An increase in thermal activity is a common precursor of volcanic eruptions and, if identified, can be used to advise local observatories to disseminate the appropriate advanced warnings. As continuously operating near-infrared (NIR) cameras are becoming more readily available at active volcanoes around the world, this investigation explores the use of identifying changes in pixel brightness in webcam imagery resulting from increased thermal incandescence. A fast, efficient, and fully automated Python algorithm has been developed with a primary focus on effective volcano monitoring and reducing overall financial costs. The algorithm includes three important tests (statistical analysis, edge detection, and Gaussian mixture model) to identify changes in activity in near-real time. The developed algorithm can be installed locally with a webcam or at a central location, with no need for additional costs. This algorithm approach was preliminarily tested on data from a permanently installed thermal infrared camera at Stromboli volcano, with a successful detection rate of 75.34%. The algorithm based methodology was further developed and applied to freely available online webcam imagery from Shiveluch volcano, with an overall accuracy of 96.0%, and a critical success index (CSI) of 76.7%. Further refinements to the algorithm were made to reduce the false alarm rate (FAR) and number of missed events, and applied to four additional image datasets at Shiveluch, Fuego, Popocatepetl, and Stromboli. The algorithm successfully identified two large eruptions at Shiveluch, between 40 minutes and 2.5 hours prior to other satellite remote sensing methods, correctly identified the beginning of a large eruption at Fuego, which corresponded with local seismic data, and successfully identified a 90-minutes window of increased activity leading to a large paroxysm event at Popocatepetl, which was describe by the local observatory as having 'little to no warning'. The algorithm underperformed at Stromboli as the images here were capture in the thermal infrared (TIR) instead of the NIR, identifying the need for further improvements to ensure the algorithm performs correctly across multiple datasets. Overall, the algorithm developed here identifies thermally incandescent activity from increases in image pixel brightness remarkably well, and would complement existing volcano observatory monitoring tools, especially in remote or financially restricted locations as the equipment and coding language used here are extremely cheap compared to many other monitoring methods.