Automatic infrasound monitoring at the Central and Eastern European Infrasound Network via Machine Learning

The Central and Eastern European Infrasound Network (CEEIN) consists of nine infrasound arrays managed by research institutes in the Czech Republic, Austria, Hungary, Ukraine, and Romania. A hybrid machine learning model was previously developed to differentiate between natural and anthropogenic sources of infrasound. This model categorizes signals from thunderstorms, activity from Mount Etna, and human-related sources, including quarry blasts, power plants, oil refineries, and the conflict in Ukraine. The dataset includes more than 100,000 labeled detections spanning from 2017 to 2024. The hybrid model combines a Convolutional Neural Network, trained on spectrograms, with a Random Forest classifier, trained on features derived from the Progressive Multi-Channel Correlation (PMCC) method, which is used for processing the raw data. The model performed well on the test data (F1 score > 0.9); however, to assess its capabilities for near-real-time monitoring, the model was retrained with randomly selected, unlabeled detections from outside the aforementioned classes. Here, we present findings from several months of automatic monitoring, assessing both single array and network processing performance.