Machine learning has been become a power tool for solving spectrum sensing problem within cognitive radio networks. Existing spectrum sensing methods typically focus on extracting features from the original data space. This paper proposes a novel spectrum sensing method called as CM-CNN-K-means, which learns both features of sensing signals in the original data space and the distribution of the features’ latent space. First, a convolution neural network (CNN) extracts correlation features of the sensing signals in a latent space using the covariance matrices (CMs). Then, a clustering algorithm (i.e., K-means) is applied for exploiting the distribution of the latent space, grouping it into clusters corresponding to the spectrum occupancy states of the primary user. The proposed approach is effective in detecting the PU's signal under noise and fading channel scenarios, as shown by experimental results.