Abstract: To address the trajectory prediction challenges caused by the heterogeneity and high-density characteristics of urban low-altitude flight targets, a hierarchical trajectory prediction framework that integrates intent inference and local planning is proposed. By combining Bayesian inference with deep learning in a hybrid inference mechanism, the probability distribution of the destination site of the flight target is estimated in real time, and local trajectory planning is performed based on the estimation results to improve the accuracy of long-term trajectory prediction. Firstly, the urban area is discretized into multiple traversable blocks, and an LSTM (long short-term memory) network is employed to learn the trajectory transition probability model from historical trajectory data. This discretization approach reduces the difficulty of directly predicting the coordinates. Secondly, the posterior probability distribution of the destination site is iteratively updated online using Bayesian theory, leveraging long-term motion intent to enhance prediction accuracy. Finally, sampling and local trajectory planning are utilized to generate the predicted trajectories that satisfy UAV (unmanned aerial vehicle) dynamic constraints. The sampling-based method reduces computational complexity and improves prediction real-time performance. Experimental results demonstrate that the average accuracy of destination site estimation reaches 0.46 (mean ± standard error: 0.175) within a range of 9 target blocks, significantly outperforming methods without prior information. The predicted trajectories cover potential maneuver paths of the target, exhibiting smaller errors, faster switching speeds, and superior performance in long-term trajectory prediction compared to the interactive multiple models (IMM) algorithm.