The deformation control of surrounding rock is a crucial aspect in the design and construction of tunnels that traverse weak geological formations. This study developed a three-dimensional numerical model of tunnel excavation that accounts for the hysteresis of initial support, based on a highway tunnel under construction in Ganzhou, China. The rationality of the numerical model was verified through on-site measurements, and orthogonal tests were conducted to analyze the impact of construction parameters on the deformation of surrounding rock. The research findings indicate that the pipe-roof can be divided into four longitudinal sections according to the distribution of stress state: the tension section ahead of the entrance casing arch, the compression section behind the tunnel face, the tension section ahead of the tunnel face, and the compression section at the rear end of the pipe-roof. The grouting reinforcement can restrict the generation and development of plastic zone in the arch of surrounding rocks, which reduce the vault settlement significantly. The primary and secondary order of the impact on tunnel vault settlement, side wall convergence and surface uplift are excavation method, pipe-roof grouting radius, initial support thickness, pipe-roof layout range.

Accurate areal rainfall estimation is essential for hydrological modeling and flood forecasting, yet method selection remains uncertain in small catchments with extremely sparse gauge networks. This event-scale study compared arithmetic mean (AM), Thiessen polygon (TP), inverse distance weighting (IDW), precipitation–elevation linear regression (ELR), multiple linear regression (MLR), and a multi-layer perceptron (MLP) in a 0.719 km² catchment monitored by three gauges. Two complementary evaluations were conducted. Station-wise leave-one-out cross-validation (LOOCV) assessed prediction at an omitted gauge, whereas a radar-referenced comparison assessed catchment-average estimates obtained from the complete gauge network. MLR produced the lowest LOOCV error (RMSE = 0.433 mm; CC = 0.777). In the radar comparison, MLP and MLR produced nearly identical RMSE values of 0.668 and 0.669 mm, respectively, and are therefore interpreted as practically similar rather than meaningfully different. All method rankings are conditional on the selected 60 h event, the three-gauge arrangement, and uncertainty in the radar reference. The findings demonstrate that station-omission performance and full-network areal estimation address different operational questions and should be considered together when selecting an interpolation method for extremely sparse networks.