Adaptive Wavelet Scattering with SegFormer for Seismic Facies Segmentation

Seismic facies classification is a fundamental task in subsurface characterization, yet existing deep learning approaches often suffer from limited interpretability and reduced sensitivity to minority facies. To address these challenges, we propose the adaptive wavelet scattering transform-SegFormer (AWST-Former) model, a hybrid framework that combines an adaptive wavelet scattering transform with a SegFormer-based segmentation backbone. In the proposed design, fixed wavelet filters are replaced by trainable Morlet wavelets whose parameters are optimized end-to-end, enabling adaptive and physically meaningful multi-scale feature extraction. A multi-order scattering feature fusion module further enhances representation quality by aggregating coefficients across different orders and scales, while the SegFormer-based segmentation network leverages both original seismic inputs and fused scattering features to achieve high-resolution facies delineation. Experiments on the F3 and Parihaka datasets demonstrate that AWST-Former consistently outperforms both standard SegFormer and non-trainable scattering baselines, with notable improvements in segmenting minority facies and complex structures. These results highlight that embedding learnable physics into neural networks provides a powerful paradigm for improving both the performance and reliability of seismic facies classification.