For visual-based unmanned aerial vehicle (UAV) forest traversal tasks, conventional reinforcement learning (RL) algorithms typically face challenges such as sluggish training dynamics, inadequate convergence, and limited generalization capabilities. To address these challenges, the input-pruned imitation reinforcement learning (IPIRL) framework is proposed in this paper, which integrates imitation learning with reinforcement learning. In the context of imitation learning, the teacher network of this paper leverages an extensive array of the UAV’s own state and environmental perception data as inputs for reinforcement learning training within a simulated environment. This process can effectively mitigate the sparse reward issue encountered in training processes that rely solely on visual perception information. Considering practical application requirements, this paper aims to train an autonomous decision-making network for UAVs, known as the student network, which takes only onboard visual information as inputs. The student network emulates the teacher network to enhance training efficiency. Ultimately, the experimental results demonstrate that IPIRL exhibits superior generalization performance.