Unmanned Aerial Systems (UAS) collect and transmit data such as live video and radar images, which have different latency and reliability requirements, over wireless links that exhibit much performance variability. In this paper, we make three contributions. First, we show through a characterization of two real-world UAS flight datasets that there is significant opportunity to optimize data transmission in UAS settings by exploiting knowledge of UAS flight paths. Second, we developed Chimera, a system that taps into this opportunity while transmitting heterogeneous data streams over UAS networks. Chimera learns a model online that relates UAS network throughput to the flight path, and combines the model with a control framework that optimizes transmissions based on long-range throughput prediction. Third, with a combination of emulation and simulation experiments using real-world flight traces, we show Chimera’s effectiveness. Specifically, Chimera reduces penalties related to dropped radar images by 72.4%−100% compared to an algorithm agnostic to flight path information, and achieves an average bitrate of 90.5% compared to an optimal scheme that knows the exact future throughput, with only a minimal increase in radar images dropped.