Downscaling Earth System Models with Deep Learning

This repo is the PyTorch codes for "Downscaling Earth System Models with Deep Learning"

Downscaling Earth System Models with Deep Learning

Overall model architecture

Geospatial Guided Attention Module

Localization Guided Augmentation Module

Usage

- Training

usage: main_srresnet.py [-h] [--channels CHANNELS] [--batchSize BATCHSIZE]
                        [--nEpochs NEPOCHS] [--lr LR] [--step STEP] [--cuda]
                        [--start-epoch START_EPOCH] [--gpus GPUS] [--position]
                        [--cutblur] [--saliency] [--piece PIECE] [--second]
                        [--first] [--r_factor R_FACTOR]
                        [--pos_rfactor POS_RFACTOR] [--pooling POOLING]

config

  -h, --help            show this help message and exit
  --channels CHANNELS   channels to be used
  --batchSize BATCHSIZE
                        training batch size
  --nEpochs NEPOCHS     number of epochs to train for
  --lr LR               Learning Rate. Default=1e-4
  --step STEP           Sets the learning rate to the initial LR decayed by
                        momentum every n epochs
  --cuda                Use cuda?
  --start-epoch START_EPOCH
                        Manual epoch number
  --threads THREADS     Number of threads for data loader to use
  --pretrained PRETRAINED
                        path to pretrained model (default: none)
  --gpus GPUS           gpu ids
  --position            Enable position encoding
  --cutblur             Enable cutblur
  --saliency            Enable saliency detection
  --piece PIECE         pieces
  --second              Apply augmentation on second channel only
  --first               Apply augmentation on first channel only
  --r_factor R_FACTOR   R_FACTOR hyperparameter
  --pos_rfactor POS_RFACTOR
                        POS_RFACTOR hyperparameter
  --pooling POOLING     mean or max

- Evaluation

usage: evaluation.py [-h] [--channel CHANNEL] [--name NAME]
                     [--checkpoint CHECKPOINT]
optional arguments:
  -h, --help            show this help message and exit
  --channel CHANNEL     number of channels to be used
  --name NAME           name of the files
  --checkpoint CHECKPOINT
                        name of the checkpoint dir

- Test Performance Measurement

usage: compare.py [-h] [--name NAME] [--filter_season FILTER_SEASON]
                  [--data DATA]
optional arguments:
  -h, --help            show this help message and exit
  --name NAME           name of the predicted file (.npy)
  --filter_season FILTER_SEASON
  --data DATA           Dataset

Data

We provide the data for our experiment. You can download the data using following link

Output

Currently, we support the output for our model.

Dataset	Output
2x	Download
4x	Download
8x	Download

Citation

@inproceedings{park2022downscaling,
  title={Downscaling Earth System Models with Deep Learning},
  author={Park, Sungwon and Singh, Karandeep and Nellikkattil, Arjun and Zeller, Elke and Mai, Tung Duong and Cha, Meeyoung},
  booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={3733--3742},
  year={2022}
}