In high-elevation or high-latitude permafrost areas, persistent subzero temperatures significantly impact the freeze–thaw durability of concrete structures. Traditional methods for studying the frost resistance of concrete in permafrost regions do not provide a complete picture for predicting properties, and new approaches are needed using, for example, machine learning algorithms. This study utilizes four machine learning models—Support Vector Machine (SVM), extreme learning machine (ELM), long short-term memory (LSTM), and radial basis function neural network (RBFNN)—to predict freeze–thaw damage factors in concrete under low and subzero temperature conservation conditions. Building on the prediction results, the optimal model is refined to develop a new machine learning model: the Sparrow Search Algorithm-optimized Extreme Learning Machine (SSA-ELM). Furthermore, the SHapley Additive exPlanations (SHAP) value analysis method is employed to interpret this model, clarifying the relationship between factors affecting the freezing resistance of concrete and freeze–thaw damage factors. In conclusion, the empirical formula for concrete freeze–thaw damage is compared and validated against the prediction results from the SSA-ELM model. The study results indicate that the SSA-ELM model offers the most accurate predictions for concrete freeze–thaw resistance compared to the SVM, ELM, LSTM, and RBFNN models. SHAP value analysis quantitatively confirms that the number of freeze–thaw cycles is the most significant input parameter affecting the freeze–thaw damage coefficient of concrete. Comparative analysis shows that the accuracy of the SSA-ELMDE prediction set is improved by 15.46%, 9.19%, 21.79%, and 11.76%, respectively, compared with the prediction results of SVM, ELM, LSTM, and RBF. This parameter positively influences the prediction results for the freeze–thaw damage coefficient. Curing humidity has the least influence on the freeze–thaw damage factor of concrete. Comparing the prediction results with empirical formulas shows that the machine learning model provides more accurate predictions. This introduces a new approach for predicting the extent of freeze–thaw damage to concrete under low and subzero temperature conservation conditions.

Leakage in urban shallow-buried non-metallic drainage pipelines changes the moisture distribution and dielectric properties of surrounding soil, producing diffuse and overlapping anomalies in GPR B-scan images that are difficult to quantify. This study proposes a GPR-based framework for leakage detection and extent estimation to support pipeline localization and excavation-oriented decision-making. Unsaturated seepage simulation, moisture-dielectric parameter conversion, and FDTD forward modeling were integrated to generate B-scan responses under different pipe diameters, leakage apertures, and development stages. An equivalent leakage radius was defined to characterize the leakage influence range, and a range regression head was embedded into the Detection Transformer framework to develop DETR-GP for joint prediction of pipe diameter category, anomaly location, and equivalent leakage radius. A multi-source dataset comprising 1,473 FDTD-generated images and 981 field-measured B-scan images was used, and the model outputs were mapped into GIS space. DETR-GP achieved an mAP(0.5) of 90.5% on the field test set for leakage anomaly detection. Because field samples lack physically measured radius labels, the equivalent leakage radius regression was quantitatively evaluated only on FDTD validation samples, yielding a mean absolute error of 0.071 m; field radius estimates are therefore interpreted as a simulation-referenced engineering indicator rather than a field-validated physical measurement. Sandbox experiments and a representative excavation case further supported the applicability of the method for rapid screening, field reinspection, and targeted excavation.