Objective. Brain–computer interface (BCI) system has developed rapidly in the past decade. And rapid serial visual presentation (RSVP) is an important BCI paradigm to detect the targets in high-speed image streams. For decoding electroencephalography (EEG) in RSVP task, the ensemble-model methods have better performance than the single-model ones. Approach. This study proposed a method based on ensemble learning to extract discriminant information of EEG. An extreme gradient boosting framework was utilized to sequentially generate the sub models, including one global spatial-temporal filter and a group of local ones. EEG was reshaped into a three-dimensional form by remapping the electrode dimension into a 2D array to learn the spatial-temporal features from real local space. Main results. A benchmark RSVP EEG dataset was utilized to evaluate the performance of the proposed method, where EEG data of 63 subjects were analyzed. Compared with several state-of-the-art methods, the spatial-temporal patterns of proposed method were more consistent with P300, and the proposed method can provide significantly better classification performance. Significance. The ensemble model in this study was end-to-end optimized, which can avoid error accumulation. The sub models optimized by gradient boosting theory can extract discriminant information complementarily and non-redundantly.