Abstract: Advance geological prediction in tunnels faces technical challenges,including strong non-stationarity of ground-penetrating radar(GPR) signals and insufficient resolution of conventional time-frequency analyses.Hence,this study proposed an improved method based on adaptive local maximum synchrosqueezing transform(LMSST).The proposed method significantly enhanced the time-frequency resolution and noise robustness of traditional LMSST through a dynamic bandwidth optimization algorithm and local extremum search strategies.Theoretical analysis and synthetic signal testing demonstrated the superior time-frequency energy concentration characteristics of the proposed method in analyzing cross-frequency modulation components.Furthermore,the proposed method was applied to the karst tunnel section of a high-speed railway in Southwest China.Combined with the GprMax forward modeling and GPR measurements,the proposed method successfully identified geological anomalies such as karst caves.Subsequent excavation verification confirmed the identification accuracy,with positional errors of anomaly boundaries below 0.3 m.Overall,the results of this study suggest the proposed method's efficiency in enhancing time-frequency resolution and substantial engineering applicability,offering reliable technical support for tunnel construction safety in karst areas.