Evaluating data augmentation techniques in the classification of multi-station seismic-volcanic signals

Volcanoes are restless giants, and before they erupt, they often whisper. These whispers are seismic signals—subtle vibrations traveling through the earth that carry the story of magma moving, gas escaping, or rocks cracking beneath the surface. For decades, scientists have relied on human experts to listen to these whispers, manually sorting through endless streams of data to identify what is happening deep underground. This process is slow, labor-intensive, and prone to human error, especially when the signals are faint or the data is overwhelming. In recent years, computers have begun to take over this listening duty, using artificial intelligence to recognize patterns in the vibrations. However, these computer brains face a significant hurdle: they need vast amounts of examples to learn, and in the world of volcanoes, rare and dangerous events are often the hardest to find in large numbers. When a computer tries to learn from a small, unbalanced set of examples, it often becomes biased, favoring the common events it sees most often and missing the rare, critical ones that signal a true danger.

A team of researchers from the Universidad Católica del Norte in Chile set out to solve this problem by teaching computers how to learn from fewer examples without losing their ability to recognize the truth. They focused on the Lascar Volcano in northern Chile, a site monitored by a network of eleven seismic stations that record the earth's tremors. The team gathered a database of over six thousand seismic events, which experts had already sorted into five distinct categories: long-period events, volcano-tectonic events, hybrid events, tectonic events, and tremors. The challenge was that these categories were not evenly represented; some types of events were common, while others were so rare that the computer struggled to learn what they looked like. To fix this, the researchers employed a strategy called data augmentation. Instead of waiting for nature to provide more rare events, they used mathematical techniques to create new, synthetic versions of the existing data. These new examples were not fake in a misleading way; they were carefully modified versions of real signals, tweaked just enough to look different while keeping their essential identity intact. The goal was to give the computer a richer, more diverse library of examples to study, hoping it would learn the fundamental rules of volcanic activity rather than just memorizing specific details.

The researchers tested several different ways to create these new examples. Some methods added a layer of static noise to the signals, similar to the hiss of a radio between stations, to see if the computer could still recognize the pattern. Others shifted the signals slightly in time or frequency, or combined parts of different signals to create new hybrids. They even used a technique inspired by biological evolution, where the computer selected the best parts of two signals and mixed them together to create a new one. To measure how well these new signals worked, the team did not just look at whether the computer got the right answer; they also measured how different the new signals were from the original ones. They needed to ensure the new data was diverse enough to teach the computer something new, but similar enough to the real world that the lessons would apply when the computer encountered a real volcano.

The results were striking. The researchers found that models trained on datasets containing only augmented data often offered superior performance in classification, with techniques like Drifting, SpecAugment, and Jittering achieving test accuracies of 97.2%, 95.4%, and 93.9%, respectively. However, the study also revealed that a model trained on a mix of real and augmented data using the "Drifting" technique achieved the highest accuracy of 98.3% when tested on real-world signals. These numbers are significant because they show that the computer learned to recognize the underlying patterns of the volcanic activity so well that it could ignore the noise and variations of the real world. The study also revealed that simply adding more data to the rare categories was not enough; the quality and diversity of that added data mattered more. The best-performing models were those that saw a wide variety of signal shapes and textures, forcing the computer to learn the core characteristics of each event type rather than relying on shortcuts.

This work suggests a new path forward for volcano monitoring. By using data augmentation, scientists can build computer models that are more robust and reliable, even when they have very few real examples of dangerous events to start with. The study demonstrated that these models could be trained on data from a single volcano and still generalize well, meaning they could potentially be applied to other volcanoes around the world. The researchers concluded that the key to success was not just having more data, but having smarter data—examples that were diverse enough to teach the computer to be flexible, yet similar enough to the real world to be useful. In the end, the computer did not just learn to count earthquakes; it learned to listen to the volcano's language with a clarity that human experts alone might struggle to achieve, offering a powerful new tool for understanding the dynamic and often dangerous evolution of volcanic centers.