Code for paper
Xie, Y., Li, Z., Bao, H., Jia, X., Xu, D., Zhou, X. and Skakun, S., 2023, June. Auto-CM: Unsupervised deep learning for satellite imagery composition and cloud masking using spatio-temporal dynamics. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 37, No. 12, pp. 14575-14583). link
@inproceedings{xie2023auto,
title={Auto-CM: Unsupervised deep learning for satellite imagery composition and cloud masking using spatio-temporal dynamics},
author={Xie, Yiqun and Li, Zhili and Bao, Han and Jia, Xiaowei and Xu, Dongkuan and Zhou, Xun and Skakun, Sergii},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={37},
number={12},
pages={14575--14583},
year={2023}
}
Cloud masking is both a fundamental and a critical task in the many Earth observation problems. This work aims to develop an unsupervised deep learning framework to detect clouds based on the differences in the spatio-temporal dynamics of the atmospheric events and land surface.
- Label-free: The training does not require any cloud labels.
- Platform-independent: The framework can be applied to any satellite platforms (e.g., PlanetScope, Landsat-8, and Sentinel-2).
The following figure shows an example of results comparing to deep clustering (DEC) and default cloud masks (Default) included in the imagery products.
We provided the codes for training and testing with PlanetScope data. Codes for other satellite platforms may be uploaded later.
AutoCM.ipynb has everything together in one notebook.
AutoCM_training.py: Training Auto-CM model for 100 epochs using PlanetScope imagery time-series data. The model does not require time-series input during testing.
AutoCM_testing.py: Generating cloud masks for new PlanetScope imagery tiles (TOA or surface reflectance) using the trained model.