Abstract: Due to the advantages of seismic data distribution in three-dimensional spatial, seismic attributes have been widely used in prediction of oil and gas, reservoir thickness and porosity. However, there are some problems like information redundancy between seismic attributes and the fuzzy relationship between seismic attributes and reservoir property parameters. Aiming at these two problems, fuzzy rough theory and machine learning are introduced into reservoir parameter prediction. By using the fuzzy rough set theory to reduce the seismic attributes, the redundant information is removed and the optimized seismic attribute combination is obtained. Then the attributes of reduction are used as input of machine learning to implemented the nonlinear mapping from seismic attribute to reservoir parameter. This method not only preserves the effective information of seismic attributes, but also avoids the difficulty of training the network model caused by too many input variables. The practical application shows that the machine learning prediction results with attribute reduction are in better agreement with the well data and have better resolution.