Streaming Deep Learning for Super Resolution on DestinE Climate Data

This repository is a demonstrator for using Streaming Deep Learning to perform Super Resolution on DestinE Climate Data. It leverages a ResNet model with 8x super-resolution capability, applied to climate datasets streamed directly from the Earth DataHub service. The repository also provides a web application for visualizing inference results and generating GeoTIFF files from both low-resolution/high-resolution outputs.

Key Features:

$\textcolor{orange}{\textsf{Model Architecture}}$: ResNet architecture designed for an 8x super-resolution task. Data Streaming: Climate data are streamed via the Earth DataHub service into a DataLoader for real-time processing.

$\textcolor{orange}{\textsf{Data}}$:

• Low Resolution (LR): Climate Digital Twin (DT) temperature at 2 meters (t2m), IFS-NEMO model, hourly data on single levels.
• Ground Truth (HR): High-Resolution (HR) Climate Digital Twin temperature at 2 meters (t2m), IFS-NEMO model, hourly data on single levels.

$\textcolor{orange}{\textsf{Data Streaming}}$: Climate data are streamed via the Earth DataHub service into a DataLoader for real-time processing without the need to download them locally.

$\textcolor{orange}{\textsf{Web Application}}$: To visualize inference results and generating GeoTIFF files from both low-resolution/high-resolution outputs.

Prerequisites

1. Install Python Download and install Python

   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
   sh Miniconda3-latest-Linux-x86_64.sh

2. Clone the repository:

   git git@github.com:tetaud-sebastien/DestinE_eXchange_SR.git
   cd DestinE_eXchange_SR

3. Install the required packages Create python environment:

   conda create --name env python==3.12

   Activate the environment

   conda activate env

   Install python package

   pip install -r requirements.txt

Install COG file generation dependencies

• gdal_translate -> Installation on Linux

Prject Structure

    .
    ├── README.md
    ├── app.py
    ├── assets
    │   └── banner.svg
    ├── auth
    │   └── cacheb-authentication.py
    ├── cfg
    │   └── config.yaml
    ├── data
    │   ├── datasets.py
    │   └── loaders.py
    ├── inference.py
    ├── models
    │   └── models.py
    ├── notebook.ipynb
    ├── train.py
    ├── trainer.py
    └── utils
        └── general.py

Destine Service Access Login

Before to train your model, please got on the main Jupyter Notebook and the following cells to setup your destine Credentials:

%%capture cap
%run auth/cacheb-authentication.py

output_1 = cap.stdout.split('}\n')
token = output_1[-1][0:-1]

from pathlib import Path
with open(Path.home() / ".netrc", "a") as fp:
    fp.write(token)

Train model

It is possible to train the model either on the following notebook: notebook.ipynb

The training script takes a configuration file as input, which parses the training parameters(TODO). You can also run the script directly using the following command:

python train.py

You can access to the Tensboard via the following typing the cli in the root directory:

tensorboard --logdir .

$\textcolor{orange}{\textsf{config file}}$: The training use a yaml configuration file. Feel free to change parameters.

Test Model with Gradio

python app.py --model_path your/model/path.pt --config_path your/training/config.json

Example:

python app.py --model_path lightning_logs/version_0/checkpoints/best-val-ssim-epoch=49-val_ssim=0.54.pt --config_path lightning_logs/version_0/config.json

Future Work

• Expand to N Parameters: Although the current setup processes only one variable (t2m), the trainer will be extended to handle multiple parameters simultaneously.
• Multiple Sources: The architecture is flexible enough to incorporate data from various sources, allowing for a multi-source approach to super-resolution tasks in climate modeling

Getting Help

Feel free to ask questions at the following email adress: sebastien.tetaud@esa.int or open a ticket.