This paper proposes a novel algorithm for distinguishing scenery information from cloud noise in the low-level and high-level detail coefficients using the wavelet decomposition. Also this paper shows approximate coefficients only containing the scenery information, and high-level detail coefficients mainly including the cloud noise and the partial scenery information. Usually cloud is brighter than the scene illumination. Therefore the appropriate brightness threshold is setup for processing high-level detail coefficients aimed at the elimination of cloud noise. Simultaneously to remove the residual cloud at the low frequency component and improve the clarity of the scenery image, the paper further decomposes the detail coefficients based on the frequency. For example, the low-level detail coefficients are decomposed further once or twice by wavelet packets. So we can remove remaining cloud decomposed effectively at the low frequency, and through assigning the appropriate weight to the detail coefficient, achieve the goal for enhancing scenery information and improving the image clarity. Considering influence of the parameter changes on the algorithm performance, we use the entropy as the criterion for choosing the optimal parameters step by step. We have demonstrated that this algorithm using the entropy as criterion is feasible. The experimental results are superior to homomorphism filtering and the Retinex algorithm in many aspects.