Hybrid machine teaching with human oversight for classification of seismograms

Tools

Sobot, Tamara and Murray, David and Stankovic, Vladimir and Stankovic, Lina (2026) Hybrid machine teaching with human oversight for classification of seismograms. Applied Soft Computing, 188. 114434. ISSN 1568-4946 (https://doi.org/10.1016/j.asoc.2025.114434)

[thumbnail of Sobot-etal-ASC-2025-Hybrid-machine-teaching-with-human-oversight-for-classification-of-seismograms]
Preview
 Text. Filename: Sobot-etal-ASC-2025-Hybrid-machine-teaching-with-human-oversight-for-classification-of-seismograms.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo
Download (2MB)| Preview

Abstract

While Artificial Intelligence (AI) is gaining popularity, two main problems limit its adoption by geoscientists for analysing and classifying large volumes of seismic recordings to monitor seismic activity associated with landslides and natural disasters. Firstly, AI often operates as a black box, whose inner workings, geoscientists do not understand and therefore are reluctant to trust its outcomes. Secondly, classification relies on learning from large labelled datasets, which are not available in practice because of the sheer volume of seismograms needing labels and the limited confidence of geoscientists in labelling low magnitude seismograms amidst anthropogenic noise. To reduce the workload on geoscientists, and accommodate the need for human agency and oversight as one of the main trustworthy AI principles, a hybrid machine teaching framework is proposed where the human teacher (domain expert) provides a few representative labels for each class based on which the machine teacher (a Siamese deep learning network) selects and labels training examples for the learner (a deep learning-based multi-classifier). The training is performed in stages, incrementally expanding training set based on provided labelled samples at each stage, resembling the way humans learn, starting with more obvious examples and progressing towards more complex ones. Leveraging the publicly available Résif Seismological Data Portal of the Super-Sauze Landslide in the Southern French Alps, characterised by low-amplitude signals, generally highly attenuated at short distances, we show that the proposed hybrid machine teaching framework effectively classifies low-magnitude earthquakes, quakes, rockfall and anthropogenic noise with F1 scores of 0.86, 0.61, 0.87 and 0.70, respectively.

ORCID iDs

Sobot, Tamara ORCID logoORCID: https://orcid.org/0000-0002-7265-8683, Murray, David ORCID logoORCID: https://orcid.org/0000-0002-5040-9862, Stankovic, Vladimir ORCID logoORCID: https://orcid.org/0000-0002-1075-2420 and Stankovic, Lina ORCID logoORCID: https://orcid.org/0000-0002-8112-1976;
CORE (COnnecting REpositories)