Abstract: The problem that the weak thrust of non-cooperative satellites, owing to its small magnitude and long duration, is highly susceptible to being submerged in observation noise, making detection extremely difficult, was prevalent in key stages such as precise orbit determination of low-Earth-orbit satellites and orbit control of deep-space probes, including midcourse corrections and near-Moon braking. To address this problem, a spatio-temporal graph convolutional network（ST-GCN） model based on temporal learning and spatial learning was proposed in this paper. By constructing the ST-GCN model, convolutional neural network（CNN）, graph neural network（GCN）, and feature fusion are introduced, and network parameter strategies such as adaptive learning rate scheduling and dynamic graph adjacency weights were adopted to improve detection accuracy and effectively solve the difficulty in detection caused by the extremely small magnitude of satellite micro-thrust. Experiments on three groups of low Earth orbit（LEO） satellite micro-thrust simulation datasets show that compared with current popular temporal models, the ST-GCN model proposed in this paper demonstrates the best overall performance in terms of both accuracy and efficiency. This method has the potential to be transferred to deep-space exploration scenarios and can provide a technical reference for the situational awareness of non-cooperative space targets.