AutoMergeNet

This repository accompanies the paper AutoMergeNet: AutoML-based M-Source Satellite Data Fusion Evaluated with Atmospheric Case Studies

Contents

The code is structured as follows:

• The automergenet folder contains a pip-installable package with code for:
  • Training
  • Model architectures
  • Data loading (for carbon monoxide and methane)
  • Experiment utility
• The config folder contains configuration files for the experiments
• The scripts folder contains scripts for:
  • Running and evaluating NAS (nas_search.py)
  • Preparing/preprocessing methane data (create_methane_dataset.py).
• The env.yaml file contains the conda packages used in the experiments

This project runs with an MIT license. Please see the LICENSE file for more details.

Important

The env.yaml file was manually compiled. If you encounter any issues with the environment, please create a new conda environment and install the packages listed in the env.yaml file. We found that for pytorch cuda support specifically, you first need to install pytorch and only then install torchvision. If you attempt to install both at the same time, it might install the CPU version instead of the GPU version.

Note on datasets

The paths to the datasets are optional arguments in the Dataset class definitions that default to "data/<dataset_name>". The datasets are not included in this repository. The datasets used in the experiments are:

• Carbon monoxide
• Methane

Important

Update the paths (data_dir and annotations_path) in the Dataset class definitions to point to the correct location of the datasets on your machine (absolute path).

How to use

1. Download the methane data
2. Process the methane data using the create_methane_dataset.py script.
3. Download the carbon monoxide data
4. Normalise the carbon monoxide data
5. Update the paths to datasets in the Dataset class definitions (multi_modal_data.py and co_data.py)
6. Update save paths in the configuration files
7. Run the experiments using the scripts in the scripts folder

Credit

Please cite the following paper if you use this code:

J. Wąsala et al., "AutoMergeNet: AutoML-based M-Source Satellite Data Fusion Evaluated with Atmospheric Case Studies," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, doi: 10.1109/JSTARS.2025.3621068.