Aerodynamic roughness (z0) reflects the extent to which surface roughness elements reduce wind erosivity, and accurate monitoring of z0 is crucial for soil wind erosion models. Spring is a period of high soil wind erosion risk in arid and semi-arid steppes in China. However, current z0 estimation models are mostly based on surface roughness features in summer and cannot represent actual spring surface conditions. Using four machine learning methods: eXtreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Random Forest (RF), and Partial Least Squares Regression (PLSR), we evaluated the accuracy of z0 estimation models constructed based on photosynthetic vegetation (PV) and non-photosynthetic vegetation (NPV) parameters in spring (April-May). The NPV-based z0 model outperformed the PV-based z0 model, with the XGBoost_NPV-based z0 achieving the highest accuracy (R2 = 0.790, RMSECV = 0.113, rRMSECV = 0.452). From 2010 to 2022, daily spring z0 of Xilingol Steppe exhibited a decreasing trend, with a multi-year average of 0.38 cm. Spatially, z0 increased from southwest to northeast, and its stability ranked from high to low as: desert steppe, meadow steppe, typical steppe, and sandy steppe. These findings demonstrate the applicability of NPV parameters for estimating spring z0 and support long-term, large-scale monitoring of soil wind erosion.