Abstract: Radar target recognition is significantly challenged by low-slow-small (LSS) characteristics of UAV targets. To address the challenges of weak features and difficulties in temporal modeling of LSS target, a multivariate temporal feature fusion-based method is proposed in this study. First, the multivariate characteristics of targets are efficiently and robustly characterized by extracting velocity, echo, and altitude features from pure temporal trajectory data. Then, the trajectories are converted into continuous temporal segments via sliding-window slicing, preserving both local motion details and global temporal correlations. In addition, long-range dependencies are captured using temporal convolutional networks (TCN). In the experiments, an accuracy of 95.13% is achieved on the test set, with state-of-the-art methods surpassed by 1.95%. Strong generalization capability is demonstrated by a cross-dataset accuracy of 92.41%. The signal-to-noise ratio (SNR) feature is proved to be the most critical contributor in ablation studies. It is demonstrated in the sliding window size comparison experiment that the highest accuracy is achieved with an 8-point sliding window.