Hand rehabilitation robot can assist the patients in completing rehabilitation exercises. Usually these rehabilitation exercises are designed according to Fugl-Meyer Assessment(FMA). Surface electromyography(sEMG) signal is the most commonly used physiological signal to identify the patient's movement intention. However, recognizing the hand gesture based on the sEMG signal is still a challenging problem due to the low amplitude and non-stationary characteristics of the sEMG signal. In this paper, eight standard hand movements in FMA are selected for the active exercises by hand rehabilitation robots. A total of 15 volunteers' sEMG signals are collected in the course of the experiment. Four time domain features, integral EMG(IEGM), root mean square(RMS), zero crossings(ZC) and energy percentage(EP), are used to identify hand gestures. A feedforward spiking neural network receives the above time domain feature data, and combines the population coding with the Spikeprop learning algorithm to realize the accurate recognition of hand gestures. The experimental results show that: (1) the spiking neural network can achieve a satisfactory classification accuracy by using only 15 neurons; (2) the classification accuracy using all four features are highest with an accuracy of 96.5%; (3) under the same number of neurons, the classification accuracy of the spiking neural network is higher than that of the multilayer perceptron, radial basis function network and support vector machine. This demonstrates the fact that spiking neural networks can achieve a satisfactory classification accuracy with a smaller network size.