Improving Land Surface Monitoring: Fusion of Landsat 8 and Sentinel-2 Using Multiscale and Attention Mechanism

Higher resolution satellite imagery is very important for various applications, especially for monitoring any conditions on a surface level. Sentinel-2 provides data at a 10-meter resolution and Landsat 8 at a 30-meter resolution. We are in a need to fuse these datasets to get higher-resolution data series. Existing fusion techniques face challenges in extracting features effectively from rapidly changing land cover areas, as they rely on simplistic processing methods for both coarse and fine images. To address these limitations, this model uses Sentinel-2’s 10-meter resolution data and auxiliary Landsat 8 data (processed to 15-meter resolution) to enhance Landsat 8’s 30-meter data to a finer 10-meter resolution. It uses multiscale and attention mechanisms to separately take features from Landsat and Sentinel images and these extracted information are then combined, and additional information is refined using residual networks .Finally, the model combines these refined features with the roughly processed outputs from our model to produce high-accuracy predictions which can be used for various applications. The results showed that this model consistently outperformed baseline models across multiple evaluation metrics, demonstrating its effectiveness in combining Sentinel-2 and Landsat 8 data to produce 10-meter resolution imagery.