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

Technical Summary: THALIA: A Global, Multi-Modal Dataset for Volcanic Activity Monitoring

Problem Statement
Monitoring volcanic activity is critical for safeguarding lives and infrastructure, yet only a small fraction of the world's volcanoes are continuously observed. While ground-based networks (e.g., GNSS) exist, many volcanoes remain unmonitored. Satellite-based Interferometric Synthetic Aperture Radar (InSAR) offers a global solution for detecting ground deformation, a reliable precursor to eruptions. However, interpreting InSAR data is complex; distinguishing true ground deformation from atmospheric propagation delays (caused by ionospheric and tropospheric variations) is a major challenge. These delays often create artifacts that mimic deformation, leading to false positives, particularly in volcanic regions with complex topography. Furthermore, the application of deep learning to automate this interpretation has been hindered by a scarcity of well-curated, multi-modal datasets. Existing datasets, such as Hephaestus, suffer from coarse spatial resolution, loss of physical pixel values due to RGB compositing, and formats unsuitable for spatiotemporal modeling.

Methodology and Dataset Construction
The authors introduce Thalia, a global, multi-modal dataset designed to address the limitations of previous resources and support deep learning-based volcanic monitoring. Thalia builds upon the Hephaestus dataset but significantly enhances its scope and structure:

• Data Structure: Thalia consists of 38 spatiotemporal datacubes covering 44 active volcanoes globally from 2014 to 2021. Unlike previous formats, Thalia utilizes a structured, multi-dimensional array format (compressed Zarr) that co-registers multiple data sources over fixed geographic regions and time, eliminating the need for additional preprocessing for alignment.
• Resolution and Content: The dataset offers a high spatial resolution of approximately 100 m × 100 m per pixel (improved from the ~333 m of Hephaestus). It integrates four primary data modalities:
  1. InSAR Products: Wrapped phase difference and coherence maps derived from Sentinel-1 imagery via the COMET-LiCSAR system.
  2. Topography: Digital Elevation Models (DEM) from the Shuttle Radar Topography Mission (SRTM).
  3. Atmospheric Variables: To address atmospheric noise, the dataset includes Total Column Water Vapor, Surface Pressure, and Vertical Integral of Temperature from the ERA5 reanalysis dataset, aligned with SAR acquisition dates.
  4. Expert Annotations: The dataset contains 19,942 annotated samples with expert-generated masks and labels. These include:
     • Deformation Masks: Binary presence of deformation.
     • Activity Type: Classification of geophysical sources (e.g., Dyke, Mogi, Sill, Spheroid, Earthquake, Unidentified).
     • Intensity Levels: Categorized by fringe count (Low, Medium, High).
     • Volcanic Phase: State of the volcano (Rest, Unrest, Rebound).
     • Noise and Quality Flags: Annotations for atmospheric fringes, glacier fringes, orbital errors, and data corruption.
     • Textual Captions: Expert descriptions of the interferograms and annotation rationale.

Benchmark and Experimental Setup
To facilitate fair evaluation, the authors established a comprehensive benchmark for two tasks: binary ground deformation classification and semantic segmentation.

• Data Split: A temporal split was applied (Training: 2014–2019; Validation: 2019; Test: 2020–2021) to preserve spatial diversity. A "leave-3-out" split was also created to test generalization to unseen volcanic regions (Galápagos, East African Rift, South Aegean).
• Time-Series Construction: Given the bi-temporal nature of InSAR, time-series were constructed by pairing a single primary acquisition date with multiple secondary dates. Sequences were fixed to a length of three to balance temporal context and data availability.
• Models: The benchmark evaluated state-of-the-art models including ResNet-50, ViT, ConvNeXt, MobileNetV3, and EfficientNetV2 for classification, and UNet, DeepLabv3, and SegFormer for segmentation. A ConvLSTM architecture was also tested to explicitly model sequential dependencies.
• Ablation Studies: Experiments compared models trained with and without atmospheric variables to assess their impact on performance.

Key Results

• Classification Performance: Models achieved strong discriminative capability, with F1-scores reaching approximately 79% in single-timestep settings. The inclusion of atmospheric variables generally improved classification performance (8 out of 10 models showed gains), suggesting that atmospheric context aids in distinguishing deformation from noise.
• Segmentation Performance: Performance was lower than classification, with F1-scores reaching ~75%. This reflects the inherent ambiguity in delineating exact deformation boundaries, even for experts. Notably, the inclusion of atmospheric variables degraded segmentation performance. The authors hypothesize that early fusion of high-resolution InSAR data with coarse atmospheric data creates a capacity bottleneck, compromising the model's ability to extract fine-scale deformation features.
• Time-Series vs. Single-Timestep: Time-series inputs using recurrent architectures (ConvLSTM) showed superior performance in segmentation tasks compared to channel-stacking approaches, likely due to better exploitation of temporal context and mitigation of atmospheric noise. However, performance in classification tasks remained comparable between the two approaches.
• Generalization: The "leave-3-out" experiment demonstrated moderate generalization capabilities. Classification F1-scores dropped by approximately 8% on unseen volcanoes, while segmentation performance dropped more significantly (20–25%), primarily due to reduced precision (oversensitivity to noise).

Significance and Claims
The authors position Thalia as a foundational resource for advancing the intersection of machine learning and Earth science.

• First of its Kind: Thalia is claimed to be the first large-scale, machine-learning-ready dataset to integrate high-resolution InSAR products, topographic data, and atmospheric variables into a structured spatiotemporal format with expert annotations.
• Addressing Limitations: By providing wrapped phase data (preserving physical interpretability) and multi-modal inputs, Thalia addresses the specific challenges of atmospheric interference and the lack of standardized benchmarks in volcanic monitoring.
• Research Directions: The dataset and benchmark highlight critical challenges, such as the difficulty of segmenting deformation in noisy environments and the trade-offs involved in fusing multi-resolution data. The authors suggest that future work should explore alternative fusion strategies (e.g., late fusion or cross-attention) to better leverage atmospheric context without sacrificing spatial precision.
• Open Science: The dataset and code are publicly available, fostering collaboration and reproducibility in the development of automated, deep-learning-based geohazard assessment tools.