An accurate detection of unused part of the spectrum is a key requirement in cognitive radio system. In this context, it is necessary to set an appropriate threshold between signal and noise. As noise power is instantaneous in nature, a periodic estimation of noise power is required. In this paper, we have proposed a combination methodology, that uses Rank order filtering in tandem with a gradient based approach. The performance of the same has been compared with two other well-known existing techniques for noise floor estimation in nonfading as well as fading scenarios, where it is found that the proposed technique outperforms the other two in terms of minimum mean square error in the estimation of spectrum occupancy.