Thalia

This is the official repository for Thalia, a global dataset for volcanic activity monitoring through InSAR imagery. Thalia builds upon Hephaestus (Bountos et al., 2022) and offers various key advantages:

• Machine Learning-ready state
• Georeferenced InSAR imagery (vs. PNG format)
• Enhanced spatial resolution at 100m GSD (vs. 333m GSD)
• Physically interpretable pixel values (vs. RGB composites)
• Zarr file format with spatial and temporal dimensions
• Additional data critical for volcanic monitoring: Digital Elevation Model (DEM), Confounding stmospheric variables

Annotations offer rich insights on the deformation type (sill, dyke, mogi, spheroid, earthquake), the intensity level (low, medium, high), the presence of non-volcanic related fringes (atmospheric, orbital, glacier) as well as the volcanic activity phase (rest, unrest, rebound). Each sample also contains a text description of the observed phenomena, ideal for language-based machine learning modelling.

You can explore a sample minicube from the dataset and investigate its structure and available annotation variables using the interactive Google Colab notebook below:

Download the Dataset

There are two ways to access the dataset:

1. Option 1: Download the full dataset and generate your own webdatasets
2. Option 2: Use pre-generated splits via Hugging Face

---

Option 1: Full Dataset Download

1. Download the dataset

   The latest version is available here: Thalia_v0

2. Export Webdatasets

   • Set the webdataset parameter to true in the configuration file.

   • If a webdataset for the timeseries length specified in the configuration file does not exist, running main.py will automatically enter webdataset creation mode.

   • Execute the command three times, once for each of the train, validation, and test splits, to generate the full dataset.

   • After the process completes, run the renaming script:

     ./webdataset_renaming.sh

   Warning: You may need to edit the bash script to specify the correct base directory and timeseries length before running.

---

Option 2: Use Pre-generated Webdatasets

The webdatasets used in the paper and for benchmarking are available via Hugging Face:
https://huggingface.co/datasets/orion-ai-lab/Thalia

Benchmark

The code in this repository implements an extensive benchmark on Thalia with a wide range of state of the art Deep Learning models. The benchmark consists of two basic tasks: image classification and semantic segmentation, each with both single-image and time-series input. Below we list the models used in the experimentation:

Classification:

• ResNet
• MobileNet v3
• EfficientNet v2
• ConvNeXt
• ViT

Segmentation:

• DeepLab v3
• UNet
• SegFormer

Data split

We consider a temporal split, using data in 01/2014-05/2019 for training, 06/2019-12/2019 for validation and 01/2020-12/2021 for testing.

How to train a model

To train and evaluate a Deep Learning model, simply run the following (with optional flags):

python main.py

Flags:

--wandb to sync results to a Wandb (Weights and Biases) project specified by configs

The model and backbone to use, as well as various training hyperparameters (e.g. batch size, number of epochs, learning rate, etc.) need to be configured in the configuration file (configs/configs.json).

Example notebook

We provide a Jupyter notebook showing how to download and explore a minicube.

Notes

Timeseries implementation

We group per primary_date and get multiple secondary_dates --> p, [s1, ..., sn].

• If len([s1, ..., sn]) > timeseries_length we choose a random subset for every epoch
• If len([s1, ..., sn]) < timeseries_length we do random duplication every epoch so that we reach the desired length

Undersampling

We export all available samples in the WebDataset format, but during training, we sample all positive examples and an equal number of negative examples per epoch.

Cropping

We apply a 512x512 crop with a random offset from the frame center, ensuring that the deformation is included in the crop if present.

Target mask

There are 4 main options for the creation of the target mask (defined as mask_target in the configuration file):

• Last: Use the mask of the last sample in the time-series
• Peak: Use a sum over all masks in the time-series
• Union: Use a union of all masks in the time-series
• All: Return all masks in the time-series (Only used for debugging)