Moisture content (MC) is vital to freeze-driedcarrots' quality and shelf life. However, traditional moisture measurement methods are time-consuming and inefficient. Therefore, this study aimed to developa rapid, nondestructive detection method utilizing terahertz time-domain spectroscopy (THz-TDS) and machine learning (ML) technology to determine the moisture content of freeze-dried carrots. The time-domain spectral data for 140 samples with varying moisture content were collected. Based on the optical parameter extraction model, the' absorption coefficient spectrum and refractive index spectrum of these samples within the terahertz frequency band were obtained. To enhance the quality of the spectral data, the acquired spectra underwent preprocessing through moving average (MA) smoothing and Savitzky-Golay (SG) smoothing. Subsequently, three feature extraction algorithms: competitive adaptive reweighting sampling (CARS), successive projection algorithm (SPA), and uninformative variable elimination (UVE), were employed to filter out the spectral variables most closely related to water content from the original spectral data. Finally, three machine learning algorithms: partial least squares regression (PLSR), back propagation artificial neural networks (BPANN), and extreme gradient boosting (XGBoost) were utilized to construct quantitative prediction models. These models were then comprehensively evaluated using model evaluation indices to determine the optimal optical parameters and the most effective algorithm combination for detecting the moisture content of freeze-dried carrots. The results indicated that the absorption coefficient spectrum accurately and effectively captured the moisture information. Pretreatment effectively reduced spectral noise, and feature extraction identified the key variables related to moisture. BPANN exhibited the best quantitative prediction performance among the machine learning algorithms tested. Specifically, the SG-CARS-BPANN model, which was based on the absorption coefficient spectrum, demonstrated the strongest predictive capability ($R_{\text{C}}^2$=0.971 2，RMSEC=0.007 3，$R_{\text{P}}^2$=0.936 6，RMSEP=0.010 7). These findings demonstrated that the combination of THz-TDS and machine learning algorithms can realize rapid and nondestructive moisture detection in freeze-dried carrots, and the established method has the potential to monitor moisture content in freeze-dried fruits and vegetables in real time during drying and storage.