Deep-learning Based Climate Downscaling

Deep-learning Based Super-resolution Model in Climate Downscaling of Precipitation Data: this is the source code of Climate Downscaling: A Deep-Learning Based Super-resolution Model of Precipitation Data with Attention Block and Skip Connections.

National Taiwan Normal University 110 semester

Advisor: Ko-Chih Wang

Author: Chia-Hao Chiang

Model Architecture

Composed of CBAMs (channel attention, spatial attention block), convolutional layers, and upsampling layer (Subpixel manner).

Purpose

To generate high-resolution precipitation data in Taiwan. Check the links below for more data information:

• Training input: ERA5 Reanalysis Data

• Label (ground truth): TCCIP Observations

Usage

setting

Set your own parameters in config.ini and put your training dataset in the folder: data/input/, which should have the naming of *yyyymmdd.npy!

Similarly, put your label data (ground truth) dataset in the folder: data/label/, which should also have namings of yyyymmdd.npy.

The dataloader.py is constructed based on the file namings and therefore they have to contain *yyyymmdd.npy suffix.

training

Direct to the root folder of main.py and then just run it: $ python main.py. The model would be saved in the model_save_dir set in config.ini, by default, at weights/.

prediction

After put the inputs you want to predict into the folder: pred_inputs/ just run pred.py: $ python pred.py. The results will be saved in the folder: pred_results. The paths of inputs and outputs for inferencing are defined as pred_input_dir and pred_save_dir in config.ini

Showcase

Prediction on the date in 2019.03.25:

Folders

• data Contains the dataset for training.

NOTICE THAT: all namings are currently hard-coded to be used in the format of *yyyymmdd.npy.

• model_utils Contains 4 .py files:

1. dataloader.py, as it's named, pulls chunks of training data paths and load one batch of training data via np.load(path).
2. MyLayersNonSeq.py contains some module or layer blocks.
3. MyModelsNonSeq.py assembles the layers or blocks from MyLayers.py to construct models.
4. train.py build training class and training related functions.

• pred_input When operating in the inference mode, that is, running the pred.py, put the input you want to feed here.

• pretrained_results When operating in the inference mode, that is, running the pred.py, the prediction is saved here.

• pretrained_weights The pre-trained weights by the author.

• results where the loss histories are stored.

• weights where the model weights are stored when running main.py to train your own.

Environment

Update the environment.yml file. Run $ conda env create -f environment.yml to install the environment (env default name = climate). This may take minutes.

• platform: Windows11
• python == 3.9.19
• tensorflow == 2.15.0 (tested on 2024.08.19)

Tested on TWCC

• platform: Windows10
• python == 3.10.3
• tensorflow == 2.12.0 (tested on 2023.07.05)