For motion recognition based on the surface electromyography ( sEMG) , reducing the channel number of sEMG electrodes could simplify the target hardware implementation, and improve the rapid response performance. However, it also has the disadvantage of coarse accuracy. In this study, we propose a sEMG recognition method by combining the single-channel sEMG decomposition and the long short-term memory (LSTM) recurrent neural networks. Firstly, the single-channel sEMG signals are decomposed into motor unit action potential trains ( MUAPTs) . Then, features are extracted from the MUAPTs, and set as inputs to train the LSTM classification model. Experiments are conducted on 6 candidates with respect to the gesture recognition scenario. Five gestures are considered as outputs of the model. Experimental results of the proposed method are extensively compared with those obtained by other three schemes, including support vector machine ( SVM) with non-decomposition data, SVM with decomposed data, and LSTM with non-decomposition data. For the sEMG of Quadratipronator, the average classification accuracy is more than 90% using the proposed method. Compared with LSTM with non-decomposition data, SVM with decomposed data, and SVM with non-decomposition data, the accuracy of the proposed method is increased by 18. 7% , 4. 17% , and 11. 53% , respectively. These results verify the efficacy of the proposed method.