Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Anomaly detection for the identification of volcanic unrest in satellite imagery

Robert Gabriel Popescu, Nantheera Anantrasirichai, Juliet Biggs

TL;DR

This paper explores the use of unsupervised deep learning on satellite data for the purpose of identifying volcanic deformation as anomalies, and presents a detector based on Patch Distribution Modeling (PaDiM), and the detection performance is enhanced with a weighted distance.

Abstract

Satellite images have the potential to detect volcanic deformation prior to eruptions, but while a vast number of images are routinely acquired, only a small percentage contain volcanic deformation events. Manual inspection could miss these anomalies, and an automatic system modelled with supervised learning requires suitably labelled datasets. To tackle these issues, this paper explores the use of unsupervised deep learning on satellite data for the purpose of identifying volcanic deformation as anomalies. Our detector is based on Patch Distribution Modeling (PaDiM), and the detection performance is enhanced with a weighted distance, assigning greater importance to features from deeper layers. Additionally, we propose a preprocessing approach to handle noisy and incomplete data points. The final framework was tested with five volcanoes, which have different deformation characteristics and its performance was compared against the supervised learning method for volcanic deformation detection.

Anomaly detection for the identification of volcanic unrest in satellite imagery

TL;DR

This paper explores the use of unsupervised deep learning on satellite data for the purpose of identifying volcanic deformation as anomalies, and presents a detector based on Patch Distribution Modeling (PaDiM), and the detection performance is enhanced with a weighted distance.

Abstract

Satellite images have the potential to detect volcanic deformation prior to eruptions, but while a vast number of images are routinely acquired, only a small percentage contain volcanic deformation events. Manual inspection could miss these anomalies, and an automatic system modelled with supervised learning requires suitably labelled datasets. To tackle these issues, this paper explores the use of unsupervised deep learning on satellite data for the purpose of identifying volcanic deformation as anomalies. Our detector is based on Patch Distribution Modeling (PaDiM), and the detection performance is enhanced with a weighted distance, assigning greater importance to features from deeper layers. Additionally, we propose a preprocessing approach to handle noisy and incomplete data points. The final framework was tested with five volcanoes, which have different deformation characteristics and its performance was compared against the supervised learning method for volcanic deformation detection.
Paper Structure (11 sections, 5 equations, 4 figures, 2 tables)

This paper contains 11 sections, 5 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Data preprocessing
  5. Patch-based feature extraction
  6. Anomaly identification
  7. Experimental results and discussion
  8. Data and case studies
  9. Detection results
  10. Study of input formats
  11. Conclusion

Figures (4)

  • Figure 1: Interferograms at Nevados Casiri, Peru ($17.47^\circ S$, $69.813^\circ W$). (a) Normal sample (no deformation) with minor atmospheric effects. (b) Normal sample with significant atmospheric effects. (c) Abnormal sample (deformation). Top row is wrapped interferograms. Bottom row is unwrapped interferograms. Each interferogram is 50 km across.
  • Figure 2: Diagram of the proposed framework
  • Figure 3: Results at Lamongan, Indonesia ($7.981^\circ S$, $113.341^\circ E$), using (a) PaDiM (original), (b) PaDiM with weighted Mahalanobis (proposed), (c) Ganomaly, (d) Diffusion, (e) Supervised learning. Top row is real deformation (anomaly). Bottom row is no deformation (normal). The brighter yellow means higher probability. Areas inside dark and bright green contours are where $P>0.5$ and $P>0.8$, respectively. Each image is 50 km across.
  • Figure 4: Example results from Nevados Casiri, Peru ($17.47^\circ S$, $69.813^\circ W$), using (left) our model and (right) supervised learning model, showing (A) deformation (anomaly), (B) artefacts (anomaly), and (C) no deformation (normal). The brighter yellow means higher probability. Areas inside dark and bright green contours are where $P>0.5$ and $P>0.8$, respectively. Each image is 50 km across.