SubTerrAInNet: A Multimodal CNN-GNN Framework for Collapse Susceptibility Prediction in Underground Mining Tunnels

Abstract

The collapse of underground mine workings remains a severe safety problem in artisanal and small-scale mining, where engineered ground support, geotechnical instrumentation, and systematic rock-mass characterization are often unavailable. Motivated by the 12 August 2026 Kassa mine-pit collapse in Barkin Ladi, Plateau State, this study presents SubTerrAInNet, a multimodal deep learning framework that integrates high-resolution tunnel imagery with low-cost geological and structural observations for segment-level collapse-susceptibility classification. The visual stream uses a transfer-learned Residual Network-50 (ResNet-50) convolutional neural network (CNN) to extract image signatures associated with cracking, seepage staining, weathering, and spalling, while a graph neural network (GNN) encodes tunnel geometry, overburden, joint orientation, water condition, crack density, and inter-segment connectivity. The dataset comprises 8,500 RGB images linked to 264 annotated tunnel segments from 15 abandoned workings in the Jos Plateau. Partitioning was performed group-wise at the physical-site level: each mine site was assigned exclusively to training, validation, or hold-out testing, so that no tunnel segment or image from a given site occurred in more than one partition. On the reported hold-out test set, SubTerrAInNet achieved 91.82% accuracy, an F1-score of 0.88, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.95. This represents a 2.82-percentage-point accuracy gain over the best conventional baseline, Extremely Randomized Trees (89.0%), and a 1.62-point gain over the strongest reported deep learning baseline, long short-term memory with attention (LSTM-Attention; 90.2%). Ablation results further show that multimodal fusion outperforms CNN-only and GNN-only variants. The final evaluation also considers collapse-class sensitivity, specificity, precision, precision- recall area under the curve (PR-AUC), probability calibration, and bootstrap confidence intervals. Gradient-weighted class activation mapping (Grad-CAM) is used to relate image-driven predictions to visible degradation precursors. SubTerrAInNet therefore provides a resource-aware framework for camera-assisted hazard screening, although prospective external validation, calibrated uncertainty estimation, and engineering review remain necessary before operational deployment.