IIGF-Net: An interpretability-driven intensity-geometry fusion framework for shield tunnel leakage segmentation in 3D point clouds

Abstract

Automated leakage detection from MLS-acquired 3D point clouds is essential for the preventive maintenance of shield tunnels. However, existing deep learning methods must reconcile the use of physically grounded handcrafted descriptors with the opaque representations learned by high-capacity networks. Without a principled means to verify whether handcrafted descriptors are effectively utilized by the trained network, descriptor design remains heuristic and the integration of domain knowledge with deep learning remains unverified. This study proposes an interpretability-driven intensity-geometry fusion network (IIGF-Net) for binary leakage segmentation in static point clouds. The framework combines nine multiscale geometric descriptors, a Fused Intensity Descriptor (FID), a two-level descriptor verification procedure, and Graph Boundary-Aware Loss (GBALoss). The verification procedure uses Gradient × Input attribution and exhaustive channel-zeroing screening across all 512 combinations of the nine geometric descriptors, followed by retraining of selected configurations to calibrate the effects of feature removal. GBALoss uses k −nearest-neighbor label relationships to emphasize leakage-background boundary points during optimization. Using an adopted DeepGCN backbone without external pretraining, IIGF-Net achieved an IoU_leakage of 87.32 ± 0.34% and an mIoU of 93.39 ± 0.17% over five runs on the validation set. IoU_leakage remained above 81% on three independent test sets. Under the tested retraining configuration, a six-descriptor subset achieved an IoU_leakage of 86.90%, compared with 87.32% for the full descriptor configuration. GBALoss increased the boundary F1-score from 72.22% to 78.47%. These results support descriptor verification as a practical basis for feature selection and indicate that the proposed components improve leakage segmentation within the evaluated tunnel datasets.