Integrating data assimilation with deep learning. Find details in this paper.
Background forecast model is U-STNx. Model training is performed through the jupyter notebook.
Key points:
- Replace Convolution2D with CConv2D custom function if circular convolution is needed. No major performace improvement
- Ensure training and autoregressive prediction uses same convolution function.
- U-STN1 +SPEnKF for regular 24 hrs DA and 1hr forecast is given in EnKF_DD_all_time.py
- U-STN1 + SPEnKF with virtual observations from U-STN12 is given in EnKF_DD_all_time_2DAv2.py
- weights and biases for U-STN12 and U-STN1 is provided in the repository.
- Change value of "lead" and run training for any "x" in U-STNx or U-NETx
- Baseline U-NET model can be trained (and tested autoregressively) with Unet_noSTN.py
Finally, the SPEnKF algorithm implementation is inspired from Tyrus Berry's presentation here. For any questions , please reach out to me at [email protected]