<p>In this article, we address the problem of spectrum sensing using uncalibrated multiple antennas when the signal transmitted by the primary user is noncircular (NC). In the proposed scheme, eigenvalues are extracted from received signal standard and complementary covariance matrices for construction of feature vector and classified in a multidimensional space for obtaining spectrum decision. This is different from existing methods where scalar decision statistic is obtained from signal covariance matrix and spectrum classification is performed using a threshold. Specifically, in the proposed technique, feature vectors are approximated using Gaussian mixture model and its parameters are obtained using expectation-maximization algorithm. The proposed method considers the feature vector to be a point of multivariate Gaussian distributions for finding posterior probability and the spectrum is classified based on the obtained likelihood value. Simulation results show that proposed scheme is robust against noise power uncertainty due to uncalibrated antennas for detection of NC signals and show significant improvement in performance compared to existing techniques.</p>