Reconstructing Historical Landsat Time Series Using a Transformer-Based Deep Learning Approach: A Case Study in the Canadian Prairies Region

The Landsat program is crucial for monitoring environmental changes, benefiting from its long record of moderate-spatial-resolution imagery, well-established calibration, and commitment to providing open-access data. However, the sparse and irregular observation intervals in Landsat time series pose challenges for applications requiring temporally consistent data. This study evaluates the performance of two deep learning models, CFC-mmRNN and a Transformer-based network, for reconstructing Landsat time series under varying data availability conditions. Accuracy is analyzed across spectral bands, seasons, and data densities to assess their effectiveness in handling irregular temporal gaps. The results indicate that the Transformer model outperforms CfC-mmRNN. While both models achieve similar accuracy in high-density cases, CfC-mmRNN’s performance declines sharper as data density decreases. In contrast, the Transformer model maintains the temporal structure of the reconstructed time series even at very low densities, with about two observations per year. These findings suggest that CfC-mmRNN remains effective for time series applications when sufficient observations are available, and it is directly applicable to forecasting tasks. The Transformer model, however, offers a more robust solution for reconstructing sparse Landsat time series, particularly in data-scarce conditions. This study underscores the importance of selecting an appropriate deep learning method to enhance Landsat time series reconstruction.