Vision-driven autonomous flight and obstacle avoidance of Unmanned Aerial Vehicles (UAVs) along complex riverine environments for tasks like rescue and surveillance requires a robust control policy, which is yet difficult to obtain due to the shortage of trainable river environment simulators and reward sparsity in such environments. To easily verify the navigation controller performance for the river following task before real-world deployment, we developed a trainable photo-realistic dynamics-free riverine simulation environment using Unity. Successful river following trajectories in the environment are manually collected and Behavior Clone (BC) is used to train an Imitation Learning (IL) agent to mimic expert behavior and generate expert guidance. Finally, a framework is proposed to train a Deep Reinforcement Learning (DRL) agent using BC expert guidance and improve the expert policy online by sampling good demonstrations produced by the DRL to increase convergence rate and policy performance. This framework is able to solve the along-river autonomous navigation task and outperform baseline RL and IL methods. The code and trainable environments are available.