<p>Seismic data, particularly pre-stack seismic data, contains rich subsurface geological information, including a variety of geological features such as sand body configurations and subsurface medium parameters like elastic properties. Seismic attributes quantitatively measure the characteristics or properties of seismic data, and most existing seismic attributes represent a synthesis of various geological information from different aspects, often leading to multiple interpretations in their geological explanations. In the context of advanced exploration targeting deep and ultra-deep layers, seismic exploration methods generally encounter significant challenges such as reduced resolution, increased uncertainty, and increased complexity in representing geological bodies. This paper introduces a novel method for deep feature extraction from seismic data, utilizing a decoupled representational learning approach. This method allows for the extraction of features that represent individual subsurface geological elements directly from complex seismic data. It employs seismic forward modeling data as training samples, encoding seismic data containing different geological features into decomposed representations with interchangeable sub-features. Through the swapping of sub-features between different samples, the method aims to generate seismic data that represents singular geological information. This iterative process updates the model parameters of the encoders and decoders, ultimately allowing the network to acquire representations that can specifically characterize physically interpretable subsurface geological elements. Empirical results from field data in Northwestern China, focusing on the prediction of thin sand body lithological boundaries, demonstrate that this new method can delineate target geological bodies more precisely than existing methods. This significantly reduces the uncertainty in seismic interpretation, offering a more accurate and reliable tool for deep and ultra-deep seismic exploration.</p>