Detecting Volcano Thermal Activity in Night Images Using Machine Learning and Computer Vision

Sergey Korolev, Igor Urmanov, Aleksei Sorokin, Olga Girina

One of the most important tasks when studying volcanic activity is to monitor their thermal radiation. To fix and assess the evolution of thermal anomalies in areas of volcanoes, specialized hardware-thermal imagers are usually used, as well as specialized instruments of modern satellite systems. The data obtained with their help contain information that makes it relatively easy to track changes in temperature and the size of a thermal anomaly. At the same time, due to the high cost of such complexes and other limitations, thermal imagers sometimes cannot be used to solve scientific problems related to the study of volcanoes. In the current paper, day/night video cameras with an infrared-cut filter are considered as an alternative to specialized tools for monitoring volcanoes’ thermal activity. In the daytime, a camera operated in the visible range, and at night the filter was removed, increasing the camera’s light sensitivity by allowing near-infrared light to hit the sensor. In that mode, a visible thermal anomaly could be registered on images, as well as other bright glows, flares, and other artifacts. The purpose of this study is to detect thermal anomalies on night images, separate them from other bright areas, and find their characteristics, which could be used for volcano activity monitoring. Using the image archive of the Sheveluch volcano as an example, this article presents the results of developing a computer algorithm that makes it possible to find and classify thermal anomalies on video frames with an accuracy of 98%. The test results are presented, along with their validation based on thermal activity data obtained from satellite systems.