README.md

Code and Models from the paper Learning from Synthetic InSAR with Vision Transformers: The case of volcanic unrest detection, IEEE Transactions on Geoscience and Remote Sensing, 2022

If you use the code or models in this repo cite our paper:

@ARTICLE{9791383,
  author={Bountos, Nikolaos Ioannis and Michail, Dimitrios and Papoutsis, Ioannis},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Learning from Synthetic InSAR with Vision Transformers: The case of volcanic unrest detection}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TGRS.2022.3180891}}

Trained models

You can download the pretrained models here.

Directory structure:

• Model zoo root

  • architecture
    • architecture checkpoint

• The available models are:

  • Convit-PL
  • DeiT-PL
  • Swin-PL
  • Swin-PL-Pseudo

Model usage example:

torch.load('swin.pt')

Train your own model

You can train a new model by executing main.py with the proper arguments. The encoder will be automatically initialized with weights pretrained on ImageNet. Example usage for a model based on the Swin Transformer:

python main.py --encoder=swin --synthetic_train_dataset=TRAIN_PATH --synthetic_val_dataset=VALIDATION_PATH --test_dataset=TEST_PATH --batch_size=40

Pseudo training

Based on the models pretrained on the synthetic dataset you can proceed with the pseudo training process by running the pseudo_training_utils.py script. Example:

python pseudo_training_utils.py --unlabeled_path=PATH_OF_UNLABELED_DATASET  --target_path=PATH_TO_STORE_PSEUDOLABELED_SAMPLES --model_root_path=PATH_OF_DOWNLOADED_MODELS --arch=ARCHITECTURE(E.G swin, deit, convit) --test_path=REAL_TEST_PATH --synthetic_validation_path=SYNTHETIC_VAL_PATH

Make sure the target path exists and has the following structure:

• Pseudo_Directory (Name is irrelevant)
  • 0
  • 1

Datasets

Training/validation datasets

The following list contains the synthetic data used in this work:

• Synthetic train dataset
• Synthetic validation Dataset

Test dataset

The test set C1, as published by [1], can be found here.

Unlabeled dataset

The unlabeled dataset used for domain adaptation can be found here.

References

[1] Bountos, Nikolaos Ioannis, et al. "Self-supervised contrastive learning for volcanic unrest detection." IEEE Geoscience and Remote Sensing Letters 19 (2021): 1-5.