Thalia: A Global, Multi-Modal Dataset for Volcanic Activity Monitoring

Paper | GitHub | Interactive Demo (Colab)

Thalia is a global, multi-modal dataset for volcanic activity monitoring through Satellite-based Interferometric Synthetic Aperture Radar (InSAR) imagery. Building upon the Hephaestus dataset, Thalia provides higher-resolution, multi-source, and multi-temporal data in a machine-learning-ready format.

Dataset Overview

Thalia consists of 38 spatiotemporal datacubes covering 7 years of volcanic activity. It integrates:

• Georeferenced InSAR imagery: 100m Ground Sample Distance (GSD) with physically interpretable pixel values.
• Topographic Data: Digital Elevation Model (DEM).
• Atmospheric Variables: Data used to account for signal delays that can mimic ground deformation.
• Rich Annotations: Includes expert labels for deformation type (sill, dyke, mogi, spheroid, earthquake), intensity level (low, medium, high), and volcanic activity phase (rest, unrest, rebound).
• Text Descriptions: Descriptive text for each sample, enabling language-based modeling.

Tasks & Benchmark

Thalia supports two primary computer vision tasks for both single-image and time-series inputs:

Image Classification

Evaluated using architectures such as:

• ResNet, MobileNet v3, EfficientNet v2, ConvNeXt, and ViT.

Semantic Segmentation

Evaluated using:

• DeepLab v3, UNet, and SegFormer.

Data Split

The dataset follows a temporal split to ensure robust evaluation:

• Training: 01/2014 – 05/2019
• Validation: 06/2019 – 12/2019
• Testing: 01/2020 – 12/2021

Citation

If you use this dataset in your research, please cite the following paper:

@article{papadopoulos2025thalia,
  title={Thalia: A Global, Multi-Modal Dataset for Volcanic Activity Monitoring},
  author={Papadopoulos, Nikolas and Bountos, Nikolaos Ioannis and Sdraka, Maria and Karavias, Andreas and Papoutsis, Ioannis},
  journal={arXiv preprint arXiv:2505.17782},
  year={2025}
}