In the context of traditional Extended Kalman Filtering (EKF) applied to ballistic target tracking, the core issue arises from the inability of a simplified point-mass ballistic model and a fixed process noise covariance matrix (Q) to adapt to the real flight dynamics, leading to ‘model mismatch’ and limiting the final filtering accuracy. This paper presents an adaptive filtering method based on Long Short-Term Memory (LSTM) networks (LSTM-EKF). This method utilizes an LSTM neural network to learn the deep relationship from target motion states to process errors from a simulation data set generated by combining various environmental factors, and thus performs dynamic estimation and adjustment of the Q matrix online. Simulation results show that this method can compensate for model uncertainties in real time and effectively overcome the limitations of traditional EKF, achieving higher filtering accuracy compared to traditional EKF.