Noise reduction usually is conducted before first arrivals picking, which could damage effective signals. This paper proposes an algorithm of blind source separation based on the second-order statistics. The method focuses on noise separation rather than noise removal. So there are no harms to effective signals. This idea might provide a new way for noise reduction. The algorithm of blind source separation based on the second-order statistics is applied to seismic data. The results show that the algorithm is effect, noises are separated and removed, and accurate first arrivals are picked up.