Addressing the challenges in determining the development of road cavities under factors such as vehicle disturbance and rainfall, as well as the difficulty in quantifying their evolution patterns, this study proposes a method for quantitatively analyzing the evolution patterns of road cavities based on ground- penetrating radar (GPR) technology combined with laboratory simulation box scenarios. The quantitative analysis method is specifically implemented as follows: firstly, GPR is used to scan the laboratory test model to collect electromagnetic wave reflection signals from underground cavities; then, the collected waveform diagrams are processed based on Matlab software. By integrating the logic of survey line arrangement, the variation in cavity cross-sectional area corresponding to each survey line is calculated, thereby enabling quantitative characterization of cavity dimensions and their development over time. Using this method, the study analyzes the influence of different moisture contents and impact loading parameters on the development behavior of road cavities within the laboratory test model. Furthermore, a sensitivity analysis is conducted to quantitatively investigate the influence of the above three factors on the development of characteristic cross-sections of road cavities. The test results show that the coefficient of variation for the calculated cavity cross-sectional area is only 1.73%, and the relative errors of all measured values are below 5%, fully validating the high reliability of the proposed method. The mean rate of change in the middle cross-section of the cavity is relatively small, with a significantly narrower range of variation, indicating that this region is most constrained by the surrounding soil and exhibits area compression under both high-energy and repeated cumulative impacts. In contrast, the change rate at the edge cross-sections is significantly higher than that at the middle cross-section; especially under high loads and high moisture content conditions, the rate of change at both the left and right boundaries exceeds 100%. Therefore, the edge cross-section can serve as a core monitoring indicator for road cavity engineering. The factors affecting road cavity development, ranked by severity, are moisture content variation, load magnitude, and number of load applications. This study provides a reliable quantitative method and theoretical basis for the analysis of road cavity development and the formulation of road maintenance strategies in complex environments.

At present, there are relatively few studies on the settlement prediction of shield tunnels crossing existing roads in parallel. To accurately predict the influence of different factors on the settlement of existing parallel roads in the shield process, this study proposed a support vector regression prediction model based on the improved particle swarm optimization (IPSO-SVR), which was applied to the surface road settlement prediction of actual subway tunnel engineering. Based on the Changchun Metro Line 6 under-crossing Feiyue Road section project, combined with the shield boring parameters, formation information and road settlement monitoring during the construction of the shield tunnel, this paper used the grid search method of libsvm to reduce the range of hyperparameters, and improved the change of inertia weight and acceleration factor in particle swarm optimization algorithm by combining the nonlinear decline strategy. Finally, IPSO-SVR prediction model was established to achieve the settlement prediction of subsequent sections in the interval. The results show that, comparing the changes of the objective function (mean square error) in the grid search method and in the conventional particle swarm optimization training, the convergence speed of the improved particle swarm optimization is greatly improved; the convergence effect of the objective function is better, and the minimum value is reduced by nearly 15%. The mean absolute error (MAE) of IPSO-SVR prediction of road settlement proposed in this paper is 0.287, the fitting coefficient R² is 0.884, and the average relative error is only 8.91%, which has better performance than back propagation (BP) neural network, support vector regression (SVR) and particle swarm optimization support vector regression (PSO-SVR) prediction models. It can be seen that IPSO-SVR has high precision for nonlinear prediction of multi-factor coupling under complex conditions, and its prediction method is feasible and generalizable. IPSO-SVR can provide a reliable basis for effective control of road settlement and is of great significance for ensuring the normal operation of roads and the safety of shield construction.