Satellite-to-ground quantum key distribution(QKD) has verified the feasibility of wide-area quantum communication networks. Towards to the future multi-users of quantum communication networks, being able to accurately and quickly predict the key rate is the core issue for quantum network. This paper proposes a new channel prediction method based on machine learning and stellar image recognition, and applies this method to the observation of the Beijing ground station. The experimental results show that the stellar image recognition accuracy rate can reach 88%, and provide the suggestion on whether to carry out a QKD experiment. In the case of the recommended channel for satellite-to-ground QKD, it is estimated that the average rate of sifted key at elevation angle of 39.5°is about 8~9 kbit/s, and the measured sifted key rate is 8.8 kbit/s. The experimental results can be used to reasonably arrange satellite-to-ground QKD tasks of multiple satellites and multiple ground stations. Moreover, this work can improve the success rate of satellite-to-ground quantum communication, avoid wasting satellite and ground station resources, and promote the practical research of satellite-based quantum communication networking.