Deep metric learning: application to lahar early warning

Deep metric learning is a branch of deep learning based on automatic learning of a similarity measure suitable for a specific task. Deep metric learning techniques have 2 or more subnetworks which learn how to extract features that will lead to the successful qualification of a data pair as similar or different. The advantage of deep metric learning is that it requires less labeled data than the other supervised ML approaches, due to being able to construct many possible pairs of data from a small dataset. We demonstrate the usefulness of deep metric learning to provide early warning for lahars (volcanic debris flows) to the area around volcano Santiaguito, Guatemala. The developed ML model shows better performance than the traditional approach based on STA/LTA, providing 7 minutes more of warning time on average.