Automatic Signal Discrimination Using Machine Learning on the Data From the Central and Eastern European Infrasound Network

Abstract

A labeled data set of 216,681 infrasound detections was compiled using data from the Central and Eastern European Infrasound Network (CEEIN). Detections associated with quarry blasts, thunderstorms, eruptions of the Etna volcano, industrial activity, and the war in Ukraine were categorized using ground truth information, such as seismic and lightning data. To establish benchmark performance, a random forest classifier and a convolutional neural network (CNN) were trained separately, achieving F1 scores of 0.8170 and 0.8248 on the test set, respectively. An ensemble model, combining both classifiers, outperformed them achieving an F1 score of 0.8773. The model, initially trained on four CEEIN arrays, was tested on data from a separate station not included in training. Although performance initially declined, transfer learning and fine-tuning of the CNN and retraining the random forest model improved the ensemble model's F1 score to 0.9056 making it a considerable step. These results represent significant progress in automatic infrasound signal classification for monitoring the atmosphere.