Deep learning-based detection of mesoscale convective systems

This ongoing research project focuses on using deep learning for feature detection of mesoscale convective systems (MCSs). These script versions were used to detect MCSs and contribute to the following manuscript:

Feng, Z., et al. (in prep). Mesoscale Convective Systems tracking Method Intercomparison (MCSMIP): Application to DYAMOND Global km-scale Simulations. Journal of Geophysical Research: Atmospheres.

Machine learning training uses PNNL provided labels of MCS objects derived from ERA5 (Contact: Zhe Feng, PNNL).

The trained model was subsequently used with DYAMOND model data.

More broadly, this project entails a collaboration between the National Center for Atmospheric Research, the Department of Atmospheric and Oceanic Science at the University of Maryland-College Park, and the Department of Earth Sciences at the University of Connecticut, to detect MCSs across a range of datasets and models.

To detect MCSs, follow the next steps:

1. Train a U-Net to detect MCSs using preprocessed ERA5 data and DYAMOND_unet_ttr.ipynb.
2. Apply the trained U-Net to detect MCSs in the DYAMOND model output using DYAMOND_inferenceNB.ipynb.
3. Assemble the files into an organized file that contains various supplemental variables using DYAMOND_inferenceNB_part2file.ipynb.
4. Visualize new MCS objects using DYAMOND_inferenceNB_part3viz.ipynb.

Then, to create MCS tracks, follow the next steps:

1. Split MCS tracks with split_tracks.py if needed for manual parallelization.
2. Filter MCSs with write_scripts.csh, which calls mcs_filtering.py, based on common MCS criteria outlined by MCSMIP protocols.
3. Combine filtering into one file with post-process.py.

For more updated versions of these scripts, go to the corresponding GitHub repositories: 

• https://github.com/mariajmolina/ML-extremes-mcs/
• https://github.com/mg-albright/post-process-dl-mcsmip/