Developing forecast models that allow for accurate and robust prediction of crop yields is helpful for early warning of food crise. In this study, we downscaled soil moisture data using Random Forest algorithms to improve regional crop yield prediction in the US Midwest. Based on high-resolution downscaled soil moisture products, precipitation data, and visible remote sensing indices, we used three mainstream machine learning algorithms, including Support Vector Machines, Random Forests, and Artificial Neural Networks to estimate soybean yields and assessed the model performance of downscaled soil moisture in predicting crop yields by comparing several variables that characterize drought. Our research shows that Random Forest model using LST from GOES as predictors showed good performance in predicting soil moisture. In the results of yield prediction, using high-resolution SM products in machine learning models outperformed coarse-resolution SM products and precipitation products, indicating using high-resolution SM products instead of precipitation has the potential to improve the performance of crop yield prediction.