Climate change has intensified extreme droughts and strong winds in arid and semi-arid regions, accelerating soil wind erosion, dust emissions, and related environmental problems. Wind is the primary driver of soil wind erosion, with both speed and cumulative duration determining the intensity of the erosion process. Accurately characterizing regional wind speed and its cumulative duration is therefore essential for understanding the spatial distribution of soil wind erosion. Previous studies have mainly relied on weather station observations; however, their limited coverage, low temporal resolution and uneven spatial distribution have introduced significant uncertainties in regional assessments. To address these limitations, long-term reanalysis wind speed datasets have been developed, though their accuracy still needs further evaluation. In this paper, wind speed data from weather stations and from the ERA5-Land reanalysis product in the wind erosion region of northern China were collected to develop a predictive model for cumulative wind speed hours above the 5 m/s threshold, based on the XGBoost method combined with SHAP (SHapley Additive exPlanations) analysis. The results showed that, compared to the cumulative wind speed hours derived from the ERA5-Land product, the model predictions more closely aligned with corresponding weather station measurements. Across different months, the explained variance for cumulative wind speed hours in the test set increased from an initial range of 10-53 % to 60-75 %, accompanied by reductions in mean absolute error (MAE) by 0.15-0.67 hand root mean square error (RMSE) by 1.11-2.14 h. The model also performed well in both cross-time and cross-location applications, explaining 72 % of the variation in cumulative wind speed hours in the validation set. This outperformed the ERA5-Land product, which accounted for only 59 % of the variation, and significantly reduced the associated prediction errors. The constructed XGBoost model was further used to simulate the cumulative wind speed hours in the study region in 2020. Its spatial simulation accuracy was comparable to that of the ordinary kriging method, with superior performance in capturing spatial variation in areas with sparse weather station coverage. The results provide reliable data on the cumulative wind speed duration required for large-scale soil wind erosion prediction and related environmental research.