Fusing multiple frequency-decomposed seismic attributes with machine learning for thickness prediction and sedimentary facies interpretation in fluvial reservoirs

Abstract

Abstract Defining the boundaries, thicknesses and sedimentary facies of fluvial reservoirs (sand bodies) is critical for predicting hydrocarbon volumes, designing schemes for petroleum exploration and development and improving oil recovery. Most reservoirs contain thick and thin sand bodies at the same intervals, while the amplitude values of seismic data usually highlight sand bodies near the 1/4 wavelength for the tuning phenomena. Hence, the application of spectral decomposition to seismic attributes and the combination of multiple frequency-decomposed (spectral-decomposed) seismic attributes have gained increasing attention for the readjustment of tuning thickness to predict sand bodies of various thicknesses. However, the popular method of red-green-blue blending is a simple linear combination of three frequency-decomposed seismic attributes that qualitatively analyzes the sand thickness without well-log interpretation. This research proposes machine learning fusion as a new nonlinear method for fusing high-, middle-, and low-frequency seismic attributes. This method uses machine learning to link well-log interpretation and multiple-frequency seismic attributes for the quantitative prediction of sand thickness, which is important for development work in a mature field. Test results of the conceptual model and the real case indicate that the predicted sand thickness after fusing multiple frequency-decomposed seismic attributes is approximately in line with the actual thickness (correlations between 80 and 90%). Combined with the coherence attribute and the red-green-blue blending results, the distributions and histories of sedimentary facies are analyzed based on the predicted sand thickness and well data. The results suggest that the proposed method can effectively readjust the tuning thickness and improve the resolution of seismic interpretation. This method is a potentially effective technique to characterize the sand thickness and sedimentary facies in other fields using a similar geological setting and dataset.