Motor imagery electroencephalography (MI-EEG), which is an important subfield of active brain-computer interface (BCI) systems, can be applied to help disabled people to consciously and directly control prosthesis or external devices, aiding them in certain daily activities. However, the low signal-to-noise ratio and spatial resolution make MI-EEG decoding a challenging task. Recently, some deep neural approaches have shown good improvements over state-of-the-art BCI methods. In this study, an end-to-end scheme that includes a multi-layer convolution neural network is constructed for an accurate spatial representation of multi-channel grouped MI-EEG signals, which is employed to extract the useful information present in a multi-channel MI signal. Then the invariant spatial representations are captured from across-subjects training for enhancing the generalization capability through a stacked sparse autoencoder framework, which is inspired by representative deep learning models. Furthermore, a quantitative experimental analysis is conducted on our private dataset and on a public BCI competition dataset. The results show the effectiveness and significance of the proposed methodology.