This paper considers the problem of mixing matrix estimation in underdetermined blind source separation (UBSS). We propose a simple and effective detection algorithm which detects the time-frequency (TF) points occupied by only a single source for each source. The detection algorithm identifies the single source points by comparing the normalized real and imaginary parts of the TF coefficient vectors of the mixed signals, which is simpler than previously reported algorithms. Then we propose a modified similarity-based robust clustering method (MSCM) to estimate the number of sources and the mixing matrix using these detected single source points. Experimental results show the efficiency of the proposed algorithm, especially in the cases where the number of sources is unknown.