CSTANet: Cross-Frame Spatio-Temporal Aggregation Network for Moving Small-Object Detection in Satellite Videos

Abstract

With the rapid advancement of satellite video technology, detecting small moving objects in satellite videos has shown substantial application potential across domains such as military surveillance, traffic monitoring, and disaster management. However, achieving robust detection remains challenging due to three key factors: 1) limited spatial resolution; 2) complex and dynamic backgrounds; and 3) low object-background contrast. To address these issues, we propose a cross-frame spatio-temporal aggregation network (CSTANet), which deeply exploits and effectively fuses spatio-temporal features to enhance detection accuracy and robustness. Extensive experiments conducted on satellite video moving object detection (MOD) datasets demonstrate that CSTANet outperforms state-of-the-art methods in multiple metrics, including recall, precision, and $F1$ -score. This study provides an effective solution for detecting small moving objects in satellite videos and is expected to advance the development of downstream remote sensing surveillance missions.