Wind erosion potential can be assessed using the Threshold Friction Velocity (TFV) of the soil, which is not always easy to measure, especially on regional and global scales. To overcome this difficulty, the spectroscopy technique can provide a useful approach in estimating the TFV as an alternative for time-consuming wind tunnel studies in the field. In this study, we evaluated the potential of Vis-NIR spectroscopy in predicting the TFV and some TFV-related soil properties using Partial Least Square Regression (PLSR) and the Support Vector Regression (SVR). We also developed a Point Spectrotransfer Function (PSTF) using Multiple Linear Regression (MLR) to predict the TFV based on diagnostic wavelengths and compared it to the derived Pedotransfer Function (PTF). For this purpose, 300 in-situ wind tunnel tests were performed in the Fars Province, Iran and the spectral reflectance of soil samples were analysed using a spectrophotometer apparatus. The 10 best key wavelengths resulting from the correlation analysis between the TFV and the spectral reflectance were 750, 1342, 1446, 1578, 1746, 1939, 2072, 2162, 2217, and 2338 nm which were mostly located in the short-wavelength infrared (SWIR) area. The derived PSTF performed better than the PTF for the TFV estimation (R2 = 0.94, RMSE = 0.71). Results of the predictive models revealed that machine learning using the SVR had a significantly (P < 0.01) higher prediction accuracy for the TFV estimation (R2 = 0.85, RMSE = 0.45, RPD = 2.50, and RPIQ = 4.06) than the PLSR (R2 = 0.68, RMSE = 1.01, RPD = 1.72, and RPIQ = 2.64). The same results were obtained for the soil moisture, clay and CaCO3 content. This study proved that reflectance spectroscopy coupled with the machine learning algorithm is a promising technique for large-scale assessment of wind erosion.