A source-free unsupervised domain adaptation method for cross-regional and cross-time crop mapping from satellite image time series

Precise and timely information about crop types plays a crucial role in various agriculture-related applications. However, crop type mapping methods often face significant challenges in cross-regional and cross-time scenarios with high discrepancies between temporal-spectral characteristics of crops from different regions and years. Unsupervised domain adaptation (UDA) methods have been employed to mitigate the problem of domain shift between the source and target domains. Since these methods require source domain data during the adaptation phase, they demand significant computational resources and data storage, especially when large labeled crop mapping source datasets are available. This leads to increased energy consumption and financial costs. To address this limitation, we developed a source-free UDA method for cross-regional and cross-time crop mapping, capable of adapting the source-pretrained models to the target datasets without requiring the source datasets. The method mitigates the domain shift problem by leveraging mutual information loss. The diversity and discriminability terms in the loss function are balanced through a novel unsupervised weighting strategy based on mean confidence scores of the predicted categories. Our experiments on mapping corn, soybean, and the class Other from Landsat image time series in the U.S. demonstrated that the adapted models using different backbone networks outperformed their non-adapted counterparts. With CNN, Transformer, and LSTM backbone networks, our adaptation method increased the macro F1 scores by 12.9%, 7.1%, and 5.8% on average in cross-time tests and by 20.1%, 12.5%, and 8.8% on average in cross-regional tests, respectively. Additionally, in an experiment covering a large study area of 450 km × 300 km, the adapted model with the CNN backbone network obtained a macro F1 score of 92.6%, outperforming its non-adapted counterpart with a macro F1 score of 89.2%. Our experiments on mapping the same classes using Sentinel-2 image times series in France demonstrated the effectiveness of our method across different countries and sensors. We also tested our method in more diverse agricultural areas in Denmark and France containing six classes. The results showed that the adapted models outperformed the non-adapted models. Moreover, in within-season experiments, the adapted models performed better than the non-adapted models in the vast majority of weeks. These results and their comparison to those obtained by the other investigated UDA methods demonstrated the efficiency of our proposed method for both end-of-season and within-season crop mapping tasks. Additionally, our study showed that the method is modular and flexible in employing various backbone networks. The code and data are available at https://github.com/Sina-Mohammadi/SFUDA-CropMapping.