Electroencephalogram (EEG) sensor data contain rich information about human emotionality. Emotion recognition based on EEG signals has attracted growing attention of researchers, especially with the fast progress of intelligent sensing technology. Numerous methods for the issue of EEG-based emotion classification have been presented in recent years. Although these methods have promoted the research development of this issue, the performance enhancement seems very slow, because the useful emotional information is quite weak compared with severe noise interferences and serious data variations. To capture the weak emotional information from the EEG signals that are distorted by various disturbances, this paper proposes a novel and effective approach, "Spatio-Temporal Field (STF)". This method extracts the Rational Asymmetry of Spectral Power features from the EEG signals at first, and then divides the feature space into the local field via the set-based discriminative measure, and finally employs the Bidirectional Long Short-Term Memory in the local field to exploit the local dynamic information for emotion classification. Experimental results have demonstrated the advantage of STF in EEG-based emotion classification by the public challenging databases, DEAP and DREAMER. STF can be regarded as an initial attempt to deal with the issue of EEG-based emotion classification from the local field perspective. We expect that the proposed method not only can provide inspirations for the further research on this issue, but may also enrich the field methodology for more general signal classification topics.