Fast and effective classification of landslide microseismicity: a machine learning perspective

The analysis of microseismic events from passive seismic data has proven invaluable for failure prediction in the context of natural hazard and landslide monitoring. Generally, microseismicity associated with fracturing processes exponentially increases prior to failure. However, passive seismic recordings commonly include a broad range of non-fracturing signals, making reliable event classification a prerequisite for near-real-time monitoring. This issue is particularly critical under data-limited monitoring conditions, such as during the early phases of slope monitoring or at minimally instrumented sites, where only a single seismic station or a very sparse sensor configuration is available. This study evaluates and compares the performance of supervised and unsupervised machine-learning approaches for the classification of landslide-related microseismic events using a challenging single-station dataset acquired over a six-month monitoring period at the Bossea Cave (NW Italy). Fracture-related microseismicity at the site exhibits marked temporal variability in spectral features, primarily driven by air-temperature fluctuations and precipitation. A supervised convolutional neural network (CNN), trained on a limited labeled dataset, is compared with an unsupervised k-means clustering approach applied to time- and frequency-domain features reduced through principal component analysis (PCA). The k-means clustering is strongly influenced by the seasonal variability of spectral features, leading to diffuse and overlapping clusters, whereas the CNN operates directly on signal spectrograms and incorporates Monte Carlo Dropout layers to quantify epistemic uncertainty through stochastic inference. Both approaches achieve satisfactory classification performance; however, the CNN demonstrates slightly higher accuracy and improved robustness under non-stationary spectral conditions. The results emphasize both the potential and the limitations of automated single-station classification workflows for passive seismic monitoring in realistic, data-limited operational scenarios.