Structure-constrained UMamba3D for seismic fault segmentation and application in the Sichuan Basin

Abstract

Automatic fault segmentation from 3D seismic volumes remains challenging because faults are sparse, thin, and easily obscured by noise, weak reflections, acquisition footprints, and complex stratigraphic discontinuities. This study proposes UMamba3D, a lightweight 3D encoder-decoder framework for seismic fault probability estimation. UMamba3D integrates a U-Net convolutional backbone, a compact Mamba-inspired bottleneck for spatial sequence mixing, and a structure-constrained learning objective. The bottleneck reshapes low-resolution volumetric features into spatial tokens and combines depth-wise 1D convolution, layer normalization, channel projection, SiLU activation, and residual fusion, thereby enhancing contextual interaction beyond local 3D convolution without introducing full-volume self-attention. The learning objective combines class-balanced binary cross-entropy, Tversky loss, and a local structure constraint to address foreground sparsity, false-positive control, and small spatial offsets around thin fault labels. Experiments were performed on the public FaultSeg3D synthetic benchmark with 200 training cubes and 20 validation cubes of size 128 × 128 × 128, followed by application to a field seismic volume from the Sichuan Basin. Controlled architecture comparisons, functional bypass analysis, and noisy-input tests characterized the bottleneck’s role in false-positive control and fault-aligned feature modulation and showed a recall-oriented robustness advantage under severe noise. Field-data comparisons with eigen coherence, Thin Likelihood, and U-Net indicate that UMamba3D produces the most geologically interpretable and noise-resistant fault candidates among the tested methods. These findings suggest that UMamba3D is a promising approach for fault probability estimation and interpretation in noisy 3D seismic volumes.