Generalized Classification of Satellite Image Time Series with Thermal Positional Encoding

Source code for Generalized Classification of Satellite Image Time Series with Thermal Positional Encoding by Joachim Nyborg, Charlotte Pelletier, and Ira Assent, presented at EarthVision 2022.

We consider the problem of generalizing crop classifiers using satellite image time series across European regions. Our proposed method, Thermal Positonal Encoding (TPE), incorporate thermal time instead of calendar time to account for temporal shifts of crop growth timelines.

The calculation of thermal time (growing degree days) can be found in dataset.py and the implementation of TPE can be found in models/ltae.py.

Requirements

• PyTorch 1.10.0
• Python 3.8.12
• Numpy 1.21.2

The TimeMatch dataset and our extension with weather data can be downloaded from Zenodo. The data classes and splits used for the paper can be found in the dataset_extensions directory.

Usage

See scripts/run_experiments.sh for examples for how to train both calendar time and thermal time model variants.

Citation

If you find the paper and/or the code useful for your work, please consider citing our paper:

@InProceedings{Nyborg_2022_CVPR,
    author    = {Nyborg, Joachim and Pelletier, Charlotte and Assent, Ira},
    title     = {Generalized Classification of Satellite Image Time Series With Thermal Positional Encoding},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {1392-1402}
}

Credits

• The code builds upon the original TimeMatch code
• The implementation of PSE+LTAE is based on the official implementation