In this paper, a trace-transform-based radial trace time frequency peak-filtering method (RT-TFPF) with temporal-spatial directions is proposed. It utilizes the similarity of data along the reflection event and computes the temporal–spatial radial directions by seeking the local maximum value of a constructed trace function. This method takes advantage of the TFPF in non-stationary signal estimation, especially with no prior knowledge. Furthermore, applying the filtering in the temporal–spatial domain results in less biased TFPF estimation. Within the framework of the trace transform, the specified trace function first calculates the centroid and then accumulates the energy of the reflected signal along the trajectory, helping to find the locally optimal filtering directions automatically. Experiments on both synthetic record and field data in North-East China demonstrates good performance–strong random noise can be attenuated, while at the same time, the estimated reflection signal is more accurate for use in interpretation.