Abstract

Deep learning models have significantly improved automatic volcano-seismic event classification but struggle with previously unseen signals due to their closed-set assumption, leading to confident misclassification of external events, as seismic sensors often capture non-volcanic movements. This study ad- dresses this limitation by integrating a simple and computationally lightweight K-Nearest Neighbors (KNN)-based Out-of-Distribution (OOD) detection mod- ule into a CNN classifier, dividing the classification problem in a trade-off between the classification of in-distribution volcanic events (ID) and the iden- tification of non-volcanic events (OOD). We explored an input representation that integrates waveforms and frequency spectrum alongside spectrograms using Class Activation Maps to evaluate their impact in learning. We found that combining waveforms with spectrograms improves ID performance as well as OOD sensitivity. Experimental results on a Nevados del Chillán Vol- canic Complex database show that our approach reaches a mean accuracy of 93.5 % for non volcanic classes (correctly classified as OOD) maintaining 84.3 % for the classification of volcanic classes (ID) compared to 73.8 % with the only-spectrogram representation. These findings demonstrate that combining multi-domain feature representations with a lightweight KNN-based OOD module raises mean OOD accuracy from 73.8 % to 93.5 % while pre- serving an 84.3 % ID F1-score, thereby improving the reliability of automated volcano-seismic monitoring and provides evidence that the approach could be incorporated into deployable, multi-volcano systems after further validation on additional sites.