Abstract
Satellite video enables continuous observation over wide areas, capturing spatiotemporal information for dynamic airborne target monitoring and aviation activity analysis. Such information facilitates aircraft localization and continuous trajectory tracking and can further support finer-grained aircraft analysis. However, publicly available satellite video datasets with in-flight aircraft trajectories remain scarce, and existing data are largely concentrated on airport-centered observations, with limited coverage of wide-area scenes due to the difficulty of data acquisition and the high cost of trajectory annotation. To address this problem, we introduce SAVAir, a satellite video dataset that combines real remote sensing backgrounds with both real and synthetic aircraft instances. SAVAir contains 1281 videos and 99,960 frames, with 1,284,563 horizontal bounding box (HBB) annotations and 159,045 oriented bounding box (OBB) labels. Moreover, the dataset provides rich information, such as visible-instance masks, aircraft configurations, structural attributes, flight phases, and land-cover descriptions. The synthetic data generation pipeline integrates 3D structural modeling, illumination and material reflectance simulation, atmospheric radiative transfer, optical degradation, and motion dynamics, together with sequence-level product-domain calibration and trajectory constraints, to maintain cross-frame appearance and motion consistency. To the best of our knowledge, the proposed pipeline is the first physics-guided synthesis approach specifically designed for aircraft in satellite videos. The synthesized targets achieve an LPIPS score of 0.1607, lower than Copy-Paste (0.1925), SinGAN (0.1943), and Stable Diffusion (0.2112), under consistent experimental settings. Representative object detectors and multi-object trackers are further evaluated under a unified benchmark protocol, while source-disjoint transfer experiments demonstrate the utility of synthetic pretraining under limited real supervision.