In-Flight Aircraft Detection in Satellite Videos: A Benchmark Dataset and Temporal Snake Mixing Network

Satellite videos provide dynamic observations over large areas, offering an important means for in-flight aircraft monitoring. However, due to the wide operational range of aircraft and limited observation windows, real data acquisition is difficult, and related studies remain scarce. Since the small scale of aircraft targets limits the information available in individual frames, existing methods for small moving object detection typically rely on multiframe fusion to enhance detection. However, background variations in satellite videos can disrupt temporal correspondences, affecting the extraction of motion information. To address the lack of data, we construct the In-Flight Aircraft Detection Dataset (IFAirDet), which consists of extensive simulation data generated under physical constraints and real satellite video data. IFAirDet covers diverse aircraft types, motion states, and scenarios. To separate target motion information from temporal background variations, we propose the Temporal Snake Mixing Network (TSMNet). Specifically, Temporal Snake Convolution (TSC) learns continuous spatial offsets across consecutive frames to adaptively adjust sampling positions, achieving temporal feature alignment and suppressing irrelevant responses. The Temporal Mixer (TMixer) further aggregates global temporal information and models the discriminative dynamic responses generated by aircraft motion. Experimental results show that TSMNet achieves an AP50 of 75.93% on IFAirDet, improving by 8.57% over the best existing method. On the real satellite video test set, it achieves an AP50 of 65.12%, demonstrating the applicability of the IFAirDet dataset and the proposed method to real observation scenarios. Experiments on the VISO and SDM-Car datasets also validate the effectiveness of TSMNet for other weak and small moving targets in satellite videos.