Distributed Acoustic Sensing automated classifiers design via transfer learning for seismology

For a decade, Distributed Acoustic Sensing (DAS) has become an exciting new tool for research in geophysics. However, the vast amount of data produced by hundreds of virtual sensors presents a challenge, and the lack of labeled data hampers the training of new automated classifiers based on machine learning methods from scratch. Automatic classifiers are essential tools for labeling observations across large, complex, and continuous datasets. They are commonly used in seismology to classify all the geophysical events recorded by seismic stations with a high degree of accuracy. Transfer learning is a machine learning technique where a model trained on one task is repurposed on a second related task. This method is particularly beneficial as it allows for the reuse of pre-existing, labeled datasets, significantly reducing the need for new data collection and annotation. This approach is a promising way to create efficient DAS classifiers without requiring a lot of resources. We built three different classifiers: the first classifies events based on their location, distinguishing local earthquakes from distant events; the second differentiates anthropogenic events from natural ones, specifically quarry blasts from earthquakes; and the third is capable of identifying all three types of events - local earthquakes, distant events, and anthropogenic quarry blasts - simultaneously. We used random forest models to train our classifiers using a labeled dataset of 14345 seismometer signals from the NOA network made for seismological research and nuclear monitoring. We studied the transferability of those classifiers to DAS data from a new NORSAR facility called NORFOX. This infrastructure consists of 8 km of cables located in Norway with a sampling frequency of 100Hz. Its multidirectional cable configuration makes it ideal for seismological surveys. We built a small test dataset of 543 signals using 20 channels equally distributed along the array. Our approach succeeded in reaching F1 score thresholds above 76%. We also showed, by fine-tuning feature selection according to their correlations between the seismic and DAS domains, that these results can be improved. In this way, we were able to increase the scores of our models from 5.89% to 11.36%. Several transfer learning approaches were also explored and discussed. Those encouraging results highlight the potential of a transfer learning approach to build new DAS classifiers. By leveraging historical seismometer catalogs, this approach facilitates the creation of DAS classifiers, thereby saving substantial time and avoiding the need for creating specialized datasets exclusively for DAS. Several strategies for improving score thresholds for future classifiers based on these transfer learning methods can be derived.