STAIR-Tr: Transformer-based Gap Filling for Multi-source Satellite Data Fusion

Satellite imagery offers consistent and scalable data for large-scale ecosystem monitoring and provides inputs for earth system modeling. However, two primary challenges hinder the applications of satellite images: (1) the presence of data gaps due to clouds or sensor malfunctions, and (2) the trade-offs between spatial and temporal resolutions of an individual satellite sensor. Multi-source satellite data fusion is widely adopted to address these challenges, such as ESTARFM and STAIR. These methods heavily rely on cloud-free images for reference and are either manually intensive or computationally expensive, potentially limiting their scalability and application to large regions. In this study, we proposed a novel transformer-based framework that addresses these challenges by efficiently filling gaps and fusing multi-source optical satellite data, STAIR-Tr (STAIR-Transformer). Our approach utilized a transformer-based model for gap filling, followed by a multi-sensor fusion model that generates high-resolution, daily, and cloud-free satellite data. Specifically, we tested STAIR-Tr for fusing Landsat-8 (16-day and 30-m) and MODIS (daily and 500-m) data at three regions in North and South America. Compared with STAIR, STAIR-Tr significantly improved the accuracy and efficiency of gap filling, with improvements in normalized Mean Absolute Error (nMAE) from 4.83% to 3.48%, the coefficient of determination (R²) from 0.763 to 0.883, an 18-fold speedup with GPU acceleration, and a 2.3-fold cost reduction. Compared with ESTARFM, STAIR-Tr overcame the limitation of requiring cloud-free images and expanded the pool of available reference images, which further enhanced the fusion process, with an increase in R² from 0.51 to 0.64 and a decrease in nMAE from 23.2% to 20.4%. STAIR-Tr improves the accuracy and efficiency of multi-source satellite data fusion using a transformer-based gap filling architecture, making it more suitable for applications requiring frequent and high-resolution satellite observations.