Objective. Recently, transfer learning (TL) and deep learning (DL) have been introduced to solve intra- and inter-subject variability problems in brain-computer interfaces (BCIs). However, current TL and DL algorithms are usually validated within a single dataset, assuming that data of the test subjects are acquired under the same condition as that of training (source) subjects. This assumption is generally violated in practice because of different acquisition systems and experimental settings across studies and datasets. Thus, the generalization ability of these algorithms needs further validations in a cross-dataset scenario, which is closer to the actual situation. This study compared the transfer performance of pre-trained deep-learning models with different preprocessing strategies in a cross-dataset scenario.Approach. This study used four publicly available motor imagery datasets, each was successively selected as a source dataset, and the others were used as target datasets. EEGNet and ShallowConvNet with four preprocessing strategies, namely channel normalization, trial normalization, Euclidean alignment, and Riemannian alignment, were trained with the source dataset. The transfer performance of pre-trained models was validated on the target datasets. This study also used adaptive batch normalization (AdaBN) for reducing interval covariate shift across datasets. This study compared the transfer performance of using the four preprocessing strategies and that of a baseline approach based on manifold embedded knowledge transfer (MEKT). This study also explored the possibility and performance of fusing MEKT and EEGNet.Main results. The results show that DL models with alignment strategies had significantly better transfer performance than the other two preprocessing strategies. As an unsupervised domain adaptation method, AdaBN could also significantly improve the transfer performance of DL models. The transfer performance of DL models that combined AdaBN and alignment strategies significantly outperformed MEKT. Moreover, the generalizability of EEGNet models that combined AdaBN and alignment strategies could be further improved via the domain adaptation step in MEKT, achieving the best generalization ability among multiple datasets (BNCI2014001: 0.788, PhysionetMI: 0.679, Weibo2014: 0.753, Cho2017: 0.650).Significance. The combination of alignment strategies and AdaBN could easily improve the generalizability of DL models without fine-tuning. This study may provide new insights into the design of transfer neural networks for BCIs by separating source and target batch normalization layers in the domain adaptation process.