The objective of this project is to develop online machine learning algorithms in order to enable online edge-caching for hybrid satellite/terrestrial backhaul networks. These algorithms allow the hybrid satellite-terrestrial network to learn the content demand and popularity distributions and optimally exploit the terrestrial and satellite resources. The project will start from developing new machine learning techniques for online content popularity estimation and using the estimated popularity to carry out optimal edge-caching in hybrid satellite-terrestrial networks. The proposed algorithms will be able to accurately estimate both local and global content popularities. For learning, we use modern Bayesian multi-armed bandit (MAB) methods which are suitable to use for highly dynamic cellular networks in which users constantly enter and leave the cells .Then, practical transmission protocols for global and local popularities under multimodal backhauling, mixture of satellite link and terrestrial fiber link, will be proposed. The proposed protocols together with the designed caching polices will potentially offload the traffic load on terrestrial backhaul link which is a big challenge in 5G networks. To move toward a more practical model, we also consider physical layer features (e.g., terrestrial and satellite wireless channels as well as power consumption) in online edge-caching design problem. Different online caching algorithms by considering physical layer features with the objectives to reduce end-users latency and network energy consumption will be defined and solved.