The affective brain-computer interface (ABCI) aims to provide a channel for emotional communication between people and external devices. Emotion electroencephalography (EEG) recognition is the most basic and key part of the ABCI system. To adaptively select the optimal combination of spatial electrodes and frequency bands to optimize the emotion EEG feature and improve the classification effectiveness, an adaptive optimized spatial-frequency differential entropy (AOSFDE) feature is proposed. We design an importance measurement method of spatial electrodes based on the relative entropy. According to the importance of various spatial electrodes, the most important spatial electrodes are selected automatically. The sparse regression algorithm is used to optimize the differential entropy features in multiple local spatial-frequency domains. The emotion EEG database (SEED) provided by Shanghai Jiao Tong University is utilized for experimental analysis. Results show that the proposed AOSFDE method can effectively improve the recognition accuracy. For 15 subjects in this dataset, the average recognition accuracy values of positive / negative, positive / neutral and neutral/ negative binary emotional classifications are 91. 8% , 93. 3% and 85. 1% , respectively. The proposed algorithm provides a new idea and method for emotion EEG recognition.