TimeMatch

Official source code of TimeMatch: Unsupervised Cross-region Adaptation by Temporal Shift Estimation by Joachim Nyborg, Charlotte Pelletier, Sébastien Lefèvre, and Ira Assent.

Requirements

Python requirements

• Python 3.9.4, PyTorch 1.8.1, and more in environment.yml.

TimeMatch dataset download

The dataset can be freely downloaded from Zenodo. The extracted size is about 78 GB.

Pre-trained models and full results

Pre-trained models and results can also be downloaded from Zenodo

Usage

Setup conda environment and activate

conda env create -f environment.yml
conda activate timematch

Download dataset and extract to /media/data/timematch_data (or set --data_root to its path for train.py).

Pre-trained models should be extracted to timematch/outputs.

Example: train model on the source domain

python train.py -e pseltae_32VNH --source denmark/32VNH/2017 --target denmark/32VNH/2017

Train TimeMatch with pre-trained model

python train.py -e timematch_32VNH_to_30TXT --source denmark/32VNH/2017 --target france/30TXT/2017 timematch --weights outputs/pseltae_32VNH

All training scripts can be found in the scripts directory.

Reference

TODO

Credits

• The implementation of PSE+LTAE is based on the official implementation
• The implementation of competitors MMD, DANN, and CDAN is based on Transfer Learning Library, and JUMBOT on its official implementation.
• The annotations used in the TimeMatch dataset from the LPIS produced by the mapping agencies IGN (France), NaturErhverstyrelsen (Denmark), and the AMA (Austria).
• The Sentinel-2 imagery were accessed from the AWS bucket managed by Sinergise