ICELEARNING

Detection of ice core particles via deep neural networks.

Data

Download the datasets folder from Zenodo. This folder contains all the train and GRIP data. The GRIP data are placed in the datasets/test/ folder.

For each training class, you find the particle.csv file and the corresponding image folder (see column imgpaths in the csv file).

The .csv files also contain the metadata of the particles. There is also a train.csv file which includes all particle.csv files merged together.

Particle	Particle .csv file	N	Example figure
Dust	datasets/dust.csv	30758
Felsic tephra	datasets/camp.csv	7625
Basaltic tephra	datasets/grim.csv	6271
Pollen C. avellana	datasets/corylus.csv	47723
Pollen Q. robur	datasets/qrob.csv	35776
Pollen Q. suber	datasets/qsub.csv	32245
Contamination/Blurry	datasets/cont.csv	11939

Create training dataset 🏋️ & train the model 🤖

python model_train.py

This code will implement the following sequence:

• Creation of train/val/test datasets. The val dataset was used for hyperparameter tuning. Both the val and test datasets consist of a 500 image/class random subset.
• Model training and validation loops. The model performance is saved as model_training_performance.csv
• The trained model is saved as saved_model/ICELEARNING_net.pth
• The model is applied to the test dataset and the Confusion Matrix will be shown and saved as confusion_matrix_test_dataset.pdf. The model_test_results.csv file contains the predictions on the test dataset, as well as the 64d-embeddings of the last FC layer of the resnet branch.
• If CFG.run_umap_test is set to True, UMAP is run and will cluster the 64d-embeddings down to 2d. The UMAP clustering will be saved as umap_test.pdf.

Check and modify all parameters and filenames in the config file CFG.py.

Model inference on GRIP ice core samples 🕵🏿

python model_test.py

This code will implement the following sequence:

• The trained model saved_model/ICELEARNING_net.pth is loaded.
• The GRIP dataset is loaded (note: 3M+ images).
• Inference loop on each particle of the GRIP dataset.
• The test/inference_on_GRIP_samples.csv final dataset is saved. This dataset will contain the particles' metadata, the model probabilities and predictions. Another similar file is saved (test/inference_on_GRIP_samples_no_metadata.csv), which includes the 64d-embeddings of the resnet FC layern but does not include the particles' metadata.

Citation

If you find this code helpful, please cite as below:

@article{maffezzoli2023,
  title={Detection of ice core particles via deep neural networks},
  author={Maffezzoli, N. and Cook, E. and van der Bilt, W. G. M. and St{\o}ren, E. N. and Festi, D. and Muthreich, F. and Seddon, A. W. R. and Burgay, F. and Baccolo, G. and Mygind, A. R. F. and Petersen, T. and Spolaor, A. and Vascon, S. and Pelillo, M. and Ferretti, P. and dos Reis, R. S. and Sim\~oes, J. C. and Ronen, Y. and Delmonte, B. and Viccaro, M. and Steffensen, J. P. and Dahl-Jensen, D. and Nisancioglu, K. H. and Barbante, C.},
  journal={The Cryosphere},
  volume={17},
  number={2},
  pages={539--565},
  year={2023},
  url = {https://tc.copernicus.org/articles/17/539/2023/},
  doi = {10.5194/tc-17-539-2023}
  publisher={Copernicus GmbH}
}

References

1. Maffezzoli, N., Cook, E., van der Bilt, W. G. M., Støren, E. N., Festi, D., Muthreich, F., Seddon, A. W. R., Burgay, F., Baccolo, G., Mygind, A. R. F., Petersen, T., Spolaor, A., Vascon, S., Pelillo, M., Ferretti, P., dos Reis, R. S., Simões, J. C., Ronen, Y., Delmonte, B., Viccaro, M., Steffensen, J. P., Dahl-Jensen, D., Nisancioglu, K. H., and Barbante, C.: Detection of ice core particles via deep neural networks, The Cryosphere, 17, 539–565, https://doi.org/10.5194/tc-17-539-2023, 2023.

2. Please see the Discussion/Review process at https://tc.copernicus.org/articles/17/539/2023/tc-17-539-2023-discussion.html

Acknowledgments

The ICELEARNING project is supported by the European Union’s Horizon 2020 Marie Skłodowska-Curie Actions (grant no. 845115).

Contact

For questions, contact niccolo.maffezzoli@unive.it