MosaicFormer: A Novel Approach to Remote Sensing Spatiotemporal Data Fusion for Lake Water Monitors

Lake water is a crucial resource in the global hydrological cycle, providing substantial freshwater resources and regulating regional climates. High-resolution remote sensing satellites, such as Landsat, provide unprecedented opportunities for the continuous monitoring of lake area changes. However, limitations imposed by revisit cycles and cloud cover often result in only a few usable images being taken per month for a single lake, restricting our understanding of daily-scale lake dynamics. Leveraging recent advancements in AI-driven remote sensing technologies, we developed an innovative deep learning algorithm, MosaicFormer, a Transformer-based model designed for spatiotemporal fusion across diverse remote sensing applications. We used it to integrate observations from MODIS and Landsat, producing seamless daily Landsat-scale images. To demonstrate its effectiveness, we applied the model to lake monitoring, showcasing its ability to reconstruct high-resolution water body dynamics with limited Landsat data. This approach combines Masked Autoencoders (MAEs) with the Swin Transformer architecture, effectively capturing latent relationships between images. Testing on public benchmarks demonstrated that our method outperforms all traditional approaches, achieving robust data fusion with an overall R2 of 0.77. A case study on lake water monitoring reveals that our method captures daily variations in the surface area of Hala Lake, providing accurate and robust results. The results indicate that our method demonstrates significant advantages and holds substantial potential for large-scale remote sensing-based environmental monitoring.