A hybrid CNN–Mamba network with interpretable mixture-of-experts routing for tunnel lining surface defects segmentation

Accurate segmentation of tunnel lining surface defects is fundamental to service condition assessment, yet existing networks suffer from limited modeling efficiency, weak cross-defect generalization, and poor interpretability. This paper develops a hybrid CNN–Mamba architecture that unifies convolutional inductive bias with the long-range dependency modeling capability of Mamba modules. An interpretable geometric-structure, region large-kernel, and photometric-frequency mixture of experts (GRPMoE) with spatial routing is further introduced, forming the proposed GRPMoE-CNN-Mamba network. Focusing on cracks, spalling, and water leakage as typical defects, three datasets, namely SCO-D, LWLO-D, and MTLD-D, are established to support systematic analysis. Model performance and routing behavior are analyzed using segmentation metrics, pixel-wise statistics, entropy, KL divergence, and feature-space visualization. Experimental results show that GRPMoE-CNN-Mamba improves mAP@0.5, F1-score, recall and mIoU by 5.48 %, 4.76 %, 7.25 % and 2.41 %, respectively. Routing analyses further reveal stable, distinct, and complementary expert contributions across defect scenarios, providing evidence of routing-level interpretability.