Shallow-buried urban road tunnels with shafts (URTS) have reduced traffic congestion in large cities. During fire scenarios, the backflows occur at those shafts far away from the fire source inhibiting smoke exhaust, but its rules have been unknown. A 400 m (length) × 12 m (width) × 5.5 m (height) physical model with 6-7 shafts over the ceiling is established using fire dynamic simulator software. The simulations are carried out after validation by both a small-scale experiment and a full-scale experiment. A total of 16 cases with 4 heat release rates (HRRs) and 4 spacing of fire source from the nearest unit shaft #1-1 (s_f-us1), are designed. Results indicate that the smoke spreading length is nearly independent of HRR but increases with s_f-us1. Ceiling smoke temperatures follow the power exponential laws, and the attenuation coefficients decrease with the increase of HRR and s_f-us1. The farther away from the fire, the more likely the occurrence of shaft backflow. A good power exponential rule of the shaft negative mass flow rate is fitted out, and values of decay coefficient b₂ range from 0.56 to 1.0. Based on dimensional analysis, a power exponential rule of the shaft dimensionless net mass flow rate is fitted for the exhaust shafts and a linear rule for the backflow shafts. The shaft neutral plane heights range from 1.4 m to 3.6 m for the exhaust shafts and 3.2 m to 5.4 m for the backflow shafts. Also, a linear rule is fitted. This study establishes the smoke backflow theory in URTS during fire scenarios and contributes to the tunnel fire protection engineering.

Hyperspectral imaging provides a novel approach for intelligent geological perception in tunnelling and underground engineering due to its high spectral resolution, nondestructive nature, and combined spectral-spatial information. However, in confined underground spaces, noise is often introduced by short exposure times, low illumination, and dust, and limited spatial resolution can cause mixed pixel effects, complicating data processing. This study presents an underground hyperspectral imaging-based mineral mapping method that achieves wall-rock visualization and semi-quantitative mineral mapping through image denoising and spectral unmixing. A spatial-spectral recurrent transformer U-Net is developed to reduce noise by leveraging spectral band correlations and nonlocal spatial-texture dependencies. A Dirichlet-based mixed pixel simulation is used to address spectral mixing, with the N-FINDR algorithm identifying endmember minerals, and the fully constrained least squares method to estimate mineral abundances. When applied to a water diversion tunnel in Shanxi, the method generates spatial distribution maps of dolomite and calcite. The experimental results confirm its effectiveness for intelligent geological logging and subsurface geological feature analysis.