Volcanic activity classification trough Self-Supervised Learning applied to satellite radiance time series

Automated early warning systems for volcanoes, capable of early recognition of any signs of impending eruption, as well as to track the evolution of different kinds of eruptive activity in near real time, are essential to assess the volcanic hazards and mitigate the associated risk. Satellite imagery offers a systematic, synoptic framework for monitoring active volcanoes in even the most isolated corners of the Earth. 

Here we have applied a machine learning technique to automatically classify a pixel in SEVIRI imagery over an active volcano, in order to detect and characterize its eruptive activity. In particular, we have discriminated against five main classes of pixels: clear sky, cloud-contaminated, ash-contaminated, SO2-contaminated and thermal anomalies.

Due to the enormous amount of data and the difficulty in labeling it, we have used self-supervised learning to study data acquired since 2004. We have selected an area of 5x5 pixels around the active volcano under study and followed the spectral radiance variation of each pixel in the twelve bands availables. 

Our technique has been applied to several active volcanoes within the SEVIRI disk, including Etna, Nyiragongo, Stromboli, Nabro, La Palma and Fogo.