Cross-sensor data reconstruction for optical remote sensing gap-filling with attention-enhanced multi-scale fusion network

• Reconstructing 10 m Sentinel-2 images using 30 m Landsat 8 data to fill gaps caused by cloud cover. • Integrating the enhanced attention mechanism substantially improves the reconstruction performance of Sentinel-2. • The complete time-series imagery effectively extracts phenological information of crops. Landsat and Sentinel-2 satellites offer high-resolution multispectral imagery widely used for applications such as vegetation phenology assessment, soil dynamics monitoring, and crop yield estimation. However, optical remote sensing data often encounter information gaps due to satellite orbital cycles and adverse atmospheric conditions, which significantly reduce image availability and complicate accurate, timely monitoring of surface change. Although integrating Landsat and Sentinel-2 can mitigate some of these challenges, the spectral differences between the two datasets necessitate spectral adjustments to minimize discrepancies in surface reflectance. Furthermore, Sentinel-2′s distinct red-edge and narrow near-infrared spectral bands play a vital role in vegetation monitoring, and their absence significantly hinders precise tracking of vegetation growth and detailed analysis. In response to these issues, we propose a deep learning framework incorporating an enhanced attention mechanism and multi-scale connection for precise remote sensing image reconstruction. This framework leverages 30-metre Landsat-8 as input to generate the 10-metre resolution Sentinel-2, harmonizing the spectral differences between the two satellites and utilizing the reconstructed images to interpolate missing pixels. The experimental findings demonstrate that the deep learning framework in this study outperforms other methodologies across various essential quantitative metrics. The integrated enhanced attention mechanism significantly improves the accuracy of image reconstruction and visual quality, particularly in restoring fine details. The assessment of transferability further validates the model’s robustness and applicability across different periods and geographic regions. Moreover, using the reconstructed images to interpolate missing data produces a denser NDVI time series, offering great potential for downstream tasks such as crop growth monitoring, identification of key phenological stages, and enabling more precise monitoring and management in agricultural applications.