The invention discloses a method and system for training a spectrum sensing model, and a spectrum sensing method and system. The spectrum sensing method comprises the following steps of acquiring a to-be-detected radio signal in real time; obtaining a frequency spectrum sampling covariance matrix of the radio signal to be detected; converting the frequency spectrum sampling covariance matrix of the to-be-detected radio signal into th egrey-scale map information of the to-be-detected radio signal; and determining a spectrum sensing result corresponding to the to-be-detected radio signal according to the grey-scale map information of the to-be-detected radio signal. According to the present invention, by extracting the characteristics from the sampling covariance matrix of the received radiosignal by using the convolutional neural network model, the detection performance deterioration of an energy detection method under a low signal-to-noise ratio and the prior information of a signal required by a cyclic stationary feature detection method are avoided, and by using the convolutional neural network for learning the characteristics of the covariance matrix, the complexity of matrix operation during characteristic extraction and the errors during threshold estimation are avoided, and finally, the requirements of real-time detection and real-time blind detection are met.