Decoding pre-movement intention of sitting and standing with high accuracy is important to build short-delay lower limb brain-computer interface (BCI) systems. Movement-related cortical potential (MRCP) is a type of EEG activity related to pre-movement patterns. This work develops a novel approach by using spatio-temporal features of MRCP, combined with kernel partial least squares (KPLS) for feature reduction. 11 healthy subjects participated in this study and performed cue-based sit-to-stand or stand-to-sit transition. The results showed that in three-class discrimination, the mean accuracy of single type of feature is 69.6% for spatial features and 69.4% for temporal features. When the spatio-temporal features were combined and used PLS for feature reduction, the detector achieved 77.4% accuracy. Furthermore, when kernel-based method KPLS was applied, the classification accuracy raised to 82%. This approach could potentially be applied in developing low-latency BCI systems for neurorehabilitation.