/BCNN4GRACE

A Bayesian Convolutional Neural Network for reconstructing GRACE TWSA signals

Primary LanguagePython

Hydrological droughts of 2017-2018 explained by the Bayesian reconstruction of GRACE(-FO) fields

Shaoxing Mo, Yulong Zhong, Ehsan Forootan, Xiaoqing Shi, Wei Feng, Xin Yin, Jichun Wu

This is a PyTorch implementation of Bayesian Convolutional Neural Network (BCNN) for reconstructing the missing GRACE(-FO) TWSA fields of 2017-2018 from hydroclimatic predictors in an image-to-image (field-to-field) regression manner. It's revised after the repo of Dr. Yinhao Zhu. BCNN can be felxiblely applied to other problems involving complex image-to-image mappings.

Dependencies

  • python 3
  • PyTorch
  • h5py
  • matplotlib
  • scipy

BCNN network architecture

Datasets

The BCNN reconstructed data are available at https://zenodo.org/record/5336992. We have also uploaded the datasets used for BCNN training to Google Drive. One can download the datasets and train BCNN to reproduce our results.

Network Training

python train_svgd.py

The parameter settings can be modified in args.py. Two network arechitectures, CBAM and RRDB, are provided. The latter is deeper and slower, but it can generally provide a slightly better performance. The CBAM and RRDB architectures are respectively employed in our JoH paper and WRR paper.

Reference GRACE(-FO) TWSA field (a), BCNN's reconstruction (b) and predictive uncertainty (c)

The BCNN-identified drought regions during the GRACE and GRACE-FO gap (July 2017-May 2018)

Citation

See Mo et al. (2022a) and Mo et al. (2022b) for more information. If you find this repo useful for your research, please consider to cite:

* Mo, S., Zhong, Y., Forootan, E., Shi, X., Feng, W., Yin, X., Wu, J. (2022a). Hydrological droughts of 
2017–2018 explained by the Bayesian reconstruction of GRACE(-FO) fields. Water Resources Research, 58, 
e2022WR031997. https://doi.org/10.1029/2022WR031997

* Mo, S., Zhong, Y., Forootan, E., Mehrnegar, N., Yin, X., Wu, J., Feng, W., Shi, X. (2022b). Bayesian 
convolutional neural networks for predicting the terrestrial water storage anomalies during GRACE and 
GRACE-FO gap. Journal of Hydrology, 604, 127244. https://doi.org/10.1016/j.jhydrol.2021.127244

Related article: Zhu, Y., & Zabaras, N. (2018). Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification. J. Comput. Phys., 366, 415-447.

Questions

Contact Shaoxing Mo (smo@nju.edu.cn) with questions or comments.