CA-DiffSeg: Cross-Attention Guided Diffusion Model for Seismic Facies Segmentation

Seismic facies analysis infers stratigraphic depositional facies by interpreting seismic reflection characteristics, which is crucial for oil and gas reservoir prediction. Deep learning has been widely applied in seismic facies segmentation due to its strong feature extraction capabilities. While deep learning methods have demonstrated the ability to capture spatial dependencies, their performance may still be challenged in complex seismic data, especially when only limited labeled data is available. Recently, diffusion models have emerged as powerful generative frameworks capable of modeling multi-scale features in seismic data. In this study, we propose a cross-attention guided diffusion model for seismic facies segmentation. First, the seismic facies segmentation task is conceptualized as a denoising problem, where the ground truth segmentation results are used as the input to the segmentation encoder, and Gaussian noise is gradually added to this input. Second, the Morlet wavelet transform is employed to decompose the seismic data into multi-scale time-frequency features, which are used as conditional information for the diffusion model to train the neural network to reconstruct the original data. Finally, a cross-attention module is introduced to fuse the conditional feature embeddings with the segmentation feature embeddings, followed by a decoder to reconstruct the segmentation results. Experiments on the F3 and Parihaka datasets demonstrate the advantages of employing diffusion models for seismic facies segmentation and validate the effectiveness of the proposed cross-attention module.