https://github.com/lzhengchun/DSGAN

a GAN for precipitation downscaling
https://github.com/lzhengchun/DSGAN

Last synced: 28 days ago
JSON representation

a GAN for precipitation downscaling

• Host: GitHub
• URL: https://github.com/lzhengchun/DSGAN
• Owner: lzhengchun
• License: other
• Created: 2020-11-19T22:54:25.000Z (over 3 years ago)
• Default Branch: main
• Last Pushed: 2022-03-07T04:50:27.000Z (over 2 years ago)
• Last Synced: 2024-02-23T04:35:07.896Z (4 months ago)
• Language: Python
• Size: 1.73 MB
• Stars: 28
• Watchers: 3
• Forks: 11
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.md

Lists

• awesome-earth-system-ml - **Fast and accurate learned multiresolution dynamical downscaling for precipitation**

README

        
![version](https://img.shields.io/badge/Version-v1.0.0-blue.svg?style=plastic)

![PyTorch](https://img.shields.io/badge/PyTorch-v1.5.0-green.svg?style=plastic)

![license](https://img.shields.io/badge/license-CC_BY--NC-red.svg?style=plastic)

[![DOI](https://zenodo.org/badge/314389271.svg)](https://zenodo.org/badge/latestdoi/314389271)

# About

This is the offical implementation of the model, a conditinal GAN for dynamic precipitation downscaling, proposed in our paper [Fast and accurate learned multiresolution dynamical downscaling for precipitation](https://gmd.copernicus.org/articles/14/6355/2021/gmd-14-6355-2021.pdf)

# Model and Implementation 

As discussed in the manuscript, we evluated 5 different model archtecture and concluded that the `Encoded-CGAN` performed the best in different cases under different metric. So, here we only opensourced the `Encoded-CGAN`. The architecture of the generator is shown below. We can consider open other implemnetaions upon request. 

![Generator Model Arch](repo-img/DSGAN-github-encoded.png)

We need to note that, the main purpose of this repo is to `show all implementation details` of our model proposed in the paper. 

This implementation is not fine-tuned to support new dataset straightforwardly, i.e., it is NOT an out-of-box solution for any downscaling.

Depends on your dataset, you may need to handcraft our hard-coded piece to make it work with yours.

# Dataset

If you want to reproduce our results, please download the dataset from [http://doi.org/10.5281/zenodo.4298978](http://doi.org/10.5281/zenodo.4298978) and unzip them (`WRF_50km_vars-mask-clip0p05-99p5.hdf5` and `WRF_precip_2005_12km-mask-clip0p05-99p5.hdf5`) to the `dataset` folder here.

If you want to try it using your own dataset, you can either edit the data.py to load dataset of your format, or prepare your h5 dataset using the same dataset name as we used in the data.py. 

We need to note that, the main purpose of this repo is to show all implementation details of our model proposed in the paper. 

This implementation is not fine-tuned to support new dataset straightforwardly, i.e., it is not an out-of-box solution for any downscaling.

Depends on your dataset, you may need to handcraft our hard-coded piece to make it work with yours.

Basically, you need to prepare your high resolution precipitation as (omit the dimension size):

```

HDF5 "WRF_precip_2005_12km-mask-clip0p05-99p5.hdf5" {

GROUP "/" {

   DATASET "rain" {

      DATATYPE  H5T_IEEE_F32LE

      DATASPACE  SIMPLE { ( 2911, 256, 512 ) / ( 2911, 256, 512 ) }

   }

}

}

```

and your low resolution precipitation and variables as (omit the dimension size):

```

HDF5 "WRF_50km_vars-mask-clip0p05-99p5.hdf5" {

GROUP "/" {

   DATASET "IWV" {

      DATATYPE  H5T_IEEE_F32LE

      DATASPACE  SIMPLE { ( 2911, 64, 128 ) / ( 2911, 64, 128 ) }

   }

   DATASET "RAIN" {

      DATATYPE  H5T_IEEE_F32LE

      DATASPACE  SIMPLE { ( 2911, 64, 128 ) / ( 2911, 64, 128 ) }

   }

   DATASET "SLP" {

      DATATYPE  H5T_IEEE_F32LE

      DATASPACE  SIMPLE { ( 2911, 64, 128 ) / ( 2911, 64, 128 ) }

   }

   DATASET "T2" {

      DATATYPE  H5T_IEEE_F32LE

      DATASPACE  SIMPLE { ( 2911, 64, 128 ) / ( 2911, 64, 128 ) }

   }

}

}

```