DeepDA

The DeepDA model, which is a GAN model for oceanic DA, comprises a generator that implements partial convolution as well as a discriminator.

Process of DeepDA

1. Training with CESM2 Large Ensemble data set (csh/train.run.csh)
2. Generate reanalysis (i.e., test) (csh/evaluate.run.csh)

Data set (data format: netCDF4)

• You can download data set here (36.2GB): now unavailable

• Put the downloaded 'data' directory into 'DeepDA'.

• The data set consists of the following:

  (1) Training set (CESM2 LE):

  - Input
    1) Traing_Dataset_Masked_CESM2_LE_Ocean_Potential_Temperature_Anomaly.nc
    2) Observation_Mask.nc
    3) Land_Mask.nc        
  - Label
    1) Traing_Dataset_CESM2_LE_Ocean_Potential_Temperature_Anomaly.nc
  

  (2) Validation set (CESM2 LE):

  - Input
    1) Validation_Dataset_Masked_CESM2_LE_Ocean_Potential_Temperature_Anomaly.nc
    2) Observation_Mask.nc
    3) Land_mask.nc
  - Label
    1) Validation_Dataset_CESM2_LE_Ocean_Potential_Temperature_Anomaly.nc
  

  (3) Test set (Met Office Hadley Centre Intergrated Ocean Database, HadIOD 1.2.0):

  - Input
    1) HadIOD_version1.2.0.0_Ocean_Potential_Anomaly.nc
    2) Observation_Mask.nc
    3) Land_Mask.nc
  

Requirement (our environment)

• OS: Ubuntu 18.04.6 LTS
• GPU: Nvidia Tesla V100
• CUDA version: 11.4
• python = 3.6
• Tensowflow = 2.2.0 (https://www.tensorflow.org/install/)
• netCDF4 = 1.5.3
• numpy = 1.19.1

License

This project is covered under the MIT License.