Reverse modulation wireless optical communication has broad application prospects in unmanned aerial vehicle, miniaturized platform and other fields, but it is easy to be affected by atmospheric turbulence, the received signal will have large low-frequency random fluctuation, and the bit error rate is high when using the hard decision method with fixed threshold. In order to improve the accuracy of signal decision, a signal decision method based on K-means algorithm is proposed. By learning the internal characteristics of the received sample points, the sample points in the window are dynamically and adaptively classified according to their similarity. The research results show that under the condition of turbulent channel, compared with the traditional hard decision method, the signal decision method based on K-means algorithm has better bit error performance. When the bit error rate is 10-6, the signal-to-noise ratio requirement is reduced by about 2 dB.