Machine learning detection of small Strombolian eruptions in noisy multi-temporal infrared images: Toward continuous monitoring
- 1University of Louisiana at Lafayette , Louisiana ,
Abstract
We show that a Convolutional Neural Network (CNN) trained on denoised multi-temporal infrared images (single images whose red, green, and blue channels hold three consecutive frames so that change between them appears as color) can detect small Strombolian eruptions that previously fell below the detection limits of conventional automated methods. Strombolian eruptions at Mount Erebus, Antarctica, range from large eruptions that eject lava throughout the crater to small eruptions that are mostly contained within the lake. We built these images from thermal imagery of the Ray lava lake and trained a CNN to classify them into three categories: “no eruption”, “small eruption”, and “large eruption”. A training and verification dataset of more than 54,000 images was built through a three-stage process, Clean → Expanded → Denoised, starting from a curated set of eruptions. We applied confident learning (Cleanlab) with five iterations of five-fold Stratified Group K-Fold cross-validation. On the held-out test set, the denoised model achieved a macro F1 of 0.9782, Matthews Correlation Coefficient = 0.9397, and macro AUC = 0.9981. To test temporal generalization, we applied the model to all December 2015 images, recorded over 10 months after the latest training images, during which the lake’s morphology changed. The network detected 49 large and 420 small eruptions. Manual verification of all detections and 2,461 stratified-sampled “no eruption” frames confirmed per-class precision of 0.98 for “large eruption” and 0.97 for “small eruption”, with a population-weighted macro F1 of 0.93.
Related
- Fuzzy Fault Detection and Band-Conditioned Data Quality Assessment for Sensor Streams: A Continuous Glucose Monitoring Case Study
- Unsupervised ensemble methods for non-wear detection from triaxial accelerometer data in children following continuous monitoring
- Zero-Cycle: First Continuous, Unsupervised, At-Home Monitoring of Spinal Motion as a Per-Cycle Substrate for Machine Learning
- Continuous monitoring of deficit irrigation in avocado across two contrasting rainfall years using sensor networks, telemetry, and machine learning
- Development and validation of an integrated wearable system for continuous cardiorespiratory monitoring using photoplethysmography and optical respiratory detection
BibTeX
@article{dye2026machine,
title = {{Machine learning detection of small Strombolian eruptions in noisy multi-temporal infrared images: Toward continuous monitoring}},
author = {Brian C Dye and Gabriele Morra},
journal = {PNAS Nexus},
year = {2026},
month = oct,
doi = {10.1093/pnasnexus/pgag359},
url = {https://academic.oup.com/pnasnexus/advance-article/doi/10.1093/pnasnexus/pgag359/8965435},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Machine learning detection of small Strombolian eruptions in noisy multi-temporal infrared images: Toward continuous monitoring
AU - Brian C Dye
AU - Gabriele Morra
PY - 2026
DA - 2026/10/22
JO - PNAS Nexus
DO - 10.1093/pnasnexus/pgag359
UR - https://academic.oup.com/pnasnexus/advance-article/doi/10.1093/pnasnexus/pgag359/8965435
AB - We show that a Convolutional Neural Network (CNN) trained on denoised multi-temporal infrared images (single images whose red, green, and blue channels hold three consecutive frames so that change between them appears as color) can detect small Strombolian eruptions that previously fell below the detection limits of conventional automated methods. Strombolian eruptions at Mount Erebus, Antarctica, range from large eruptions that eject lava throughout the crater to small eruptions that are mostly contained within the lake. We built these images from thermal imagery of the Ray lava lake and trained a CNN to classify them into three categories: “no eruption”, “small eruption”, and “large eruption”. A training and verification dataset of more than 54,000 images was built through a three-stage process, Clean → Expanded → Denoised, starting from a curated set of eruptions. We applied confident learning (Cleanlab) with five iterations of five-fold Stratified Group K-Fold cross-validation. On the held-out test set, the denoised model achieved a macro F1 of 0.9782, Matthews Correlation Coefficient = 0.9397, and macro AUC = 0.9981. To test temporal generalization, we applied the model to all December 2015 images, recorded over 10 months after the latest training images, during which the lake’s morphology changed. The network detected 49 large and 420 small eruptions. Manual verification of all detections and 2,461 stratified-sampled “no eruption” frames confirmed per-class precision of 0.98 for “large eruption” and 0.97 for “small eruption”, with a population-weighted macro F1 of 0.93.
SN - 2752-6542
ER -