Abstract: Rapid serial visual presentation (RSVP) is a brain-computer interface (BCI) paradigm based on event-related potential (ERP) detection. By decoding and classifying electroencephalogram (EEG) signals, this technology can be widely utilized in target search and interactive control tasks. Due to the behavior of ERP in strong variability and low signal-to-noise ratio (SNR), the distribution of spatiotemporal information varies greatly for classification reflected in the cerebral cortex for different subjects. And, the performance of traditional single-trial classification algorithms based on CSP or LDA is unstable for different datasets, the robustness of classification models is limited across datasets. In order to improve the decoding performance of RSVP-BCI, two spatiotemporal filters were designed and optimized by alternating iteration for feature extraction, and a spatiotemporal analysis for ERP extraction (STAEE) algorithm was proposed based on frequency-time-space domain perspectives. The STAEE algorithm was arranged to be consisted of a filter-bank module, a time-window decomposition module, a spatiotemporal filtering module and a region of interest (ROI) selection module. In two classification tasks of public RSVP dataset, the proposed STAEE algorithm can obtained higher area under curve (AUC) values than the four benchmark algorithms, including hierarchical discriminant component analysis (HDCA), common spatial pattern (CSP), filter bank common spatial pattern (FBCSP) and space-time discriminant analysis (STDA). The results show that the STAEE algorithm can effectively overcome the variability of ERP distribution across different datasets, and improve the classification performance of RSVP-BCI system.