SSHMamba: A SAM-Assisted Scattering-Guided Hybrid Selective Mamba Network for PolSAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) image classification requires the joint use of polarimetric scattering information and spatial context. However, aggregating or propagating information across locations with different scattering characteristics may mix features from dissimilar regions, and recurrent state propagation in Mamba can extend this influence along scan paths. Class-agnostic regions generated by the Segment Anything Model (SAM) from Pauli pseudo-RGB images may also span multiple land-cover classes, which may reduce the reliability of direct regional consistency constraints. This study proposes SSHMamba, a SAM-assisted scattering-guided hybrid selective Mamba network. It constructs a seven-channel scattering prior from the Pauli components, span, and H/A/α parameters. The local branch combines center-outward spiral rearrangement, one-dimensional convolution, and neighborhood weighting based on spatial distance and scattering similarity. In the long-range branch, the scattering prior conditions the state-space parameters, while scattering similarity and a learnable gate regulate the contribution of preceding states. Multidirectional outputs are aggregated using pixel-wise direction weights, and local and long-range features are fused using scattering-prior-driven channel-wise coefficients. During training, SAM candidate regions are screened using scattering homogeneity, prediction reliability, and region area before a soft regional consistency constraint is applied; SAM is not used during inference. Across three independent training runs under the same fixed data split, SSHMamba achieves 99.12 ± 0.04% OA, 99.13 ± 0.04% AA, and 0.9904 ± 0.0005 Kappa on Flevoland, and 99.42 ± 0.02% OA, 97.52 ± 0.03% AA, and 0.9909 ± 0.0004 Kappa on San Francisco. In the fixed reference runs, ECP-Mamba yields a slightly higher OA on Flevoland (99.30% versus 99.16%), whereas SSHMamba yields a higher OA on San Francisco (99.44% versus 99.20%). On Flevoland, removing the complete scattering-guided selective state propagation mechanism reduces OA from 99.16% to 96.48%. The SAM ablation shows that direct consistency over raw SAM regions yields 97.12% OA, candidate-region reliability screening increases it to 98.72%, and the full reliability-controlled SAM-assisted strategy reaches 99.16%. On the DLR E-SAR Oberpfaffenhofen dataset, SSHMamba achieves 99.29 ± 0.07% OA, 99.15 ± 0.04% AA, and 0.9889 ± 0.0006 Kappa across three runs, providing additional cross-sensor and cross-scene evidence.