Effective and reliable methods of origin certification are essential for protecting high-value Chinese medicinal materials (e.g geo-authentic Chinese medicinal materials, geographical indication products, etc.) from designated regions. As a famous traditional Chinese medicine and a geo-authentic Chinese medicinal material produced in Sichuan Province, Aconiti Lateralis Radix Praeparata (Fuzi) has a remarkable curative effect and wide clinical application is in great demand in domestic and international markets. The efficacy and price of the Fuzi of different origins vary, and it is difficult for the public to identify them through traditional experience accurately. Mass spectrometry-based on plant metabolomics is a tedious and lengthy test sample preparation process, complicated operation, long detection time, and low reproducibility. Near-infrared (NIR) spectroscopy, a mature, fast and nondestructive detection technique was integrated with machine learning to bring new ways for online quality supervision and control of Chinese medicinal materials. Therefore, a non-destructive identification model based on NIR spectroscopy combined with a random forest (RF) algorithm was developed for different origins of Fuzi. A total of 255 samples of Fuzi were collected from the major cultivation regions of Sichuan, Shaanxi and Yunnan, and the diffuse reflectance spectral information of all samples was obtained using Fourier transform NIR spectroscopy. Single and combined spectral preprocessing methods are used to eliminate multiple interferences in the spectra, and the best preprocessing method is screened and used as an input indicator to build an RF model. The comprehensive performance of the RF model was evaluated using sensitivity, specificity and balanced accuracy. The results showed that Savitzky-Golay 11-point smoothing combined with multivariate scattering correction was the best preprocessing method.Using only the full wavelength data, the prediction accuracy of the RF model for the three groups of provincial samples was also checked over 90%, and the prediction accuracy after preprocessing reached 98.39%. For the city/county level samples, the RF model also had the excellent discriminative ability, greater than 75% accuracy. The RF model achieved 100% recognition rate for samples from cultivation areas around the traditional production areas. The top 100 feature wave numbers were filtered out, and the model was re-optimized, and the recognition accuracy of the model for each city/county level region was over 85%, especially for some samples from the highlands was significantly improved. In this study, an environment-friendly traceability strategy with faster analysis, less sample loss and higher precision was adopted, providing a new model for the rapid and efficient identification of Fuzi of different origins and a reference for the subsequent identification and traceability of Fuzi and its related processed products.