Abstract: The seismic frequency enhancement processing method based on multi-layer residual network combines high-frequency well logging information with seismic data through an intelligent network. This method effectively improves vertical resolution while maintaining lateral continuity, facilitating the identification of thin reservoirs. In the AMH area, seismic data processed by conventional techniques enable only the identification of carbonates thicker than 30 m. The seismic frequency enhancement processing method based on multi-layer residual network was proposed and applied in this area. First, training was performed using a multi-layer residual network, a deep learning network with near-wellbore seismic amplitude as training data and the relative wave impedance data from well logging curves as label data. A predictive model for relative wave impedance curve was obtained. Using seismic data as input, the deep network model calculated relative wave impedance data, and the reflection coefficient based on the frequency-enhanced seismic data. According to the geological conditions of the target area, appropriate wide-band wavelets were selected convolved with the reflection coefficient to obtain frequency-enhanced seismic data. The inversion results based on the frequency-enhanced seismic data are of high resolution vertically, and identifiable and traceable laterally, and well matched the target beds. They can identify the thin beds in the AMH area. The seismic frequency enhancement processing method based on multi-layer residual network together with corresponding high-resolution model inversion can identify beds thicker than 10 m in the AMH area. This method can effectively identify thin beds that can’t be identified using low-resolution seismic data, and improves the accuracy in predicting thin beds. It is referential for identifying similar thin beds.

Key words: carbonate, seismic data, frequency enhancement processing, thin bed, multi-layer residual network, relative wave impedance, high-resolution inversion, deep learning