Canopy closure is a critical indicator reflecting forest structure, biodiversity, and ecological balance. This study proposes an estimation method integrating U-Net segmentation with machine learning, significantly improving accuracy through multi-source remote sensing data and feature selection. Covering eight U.S. continental states, the study utilized 13,000 stratified samples equally split for model training and validation. Four states were used to train models based on XGBoost, random forest (RF), LightGBM, and support vector machine (SVM), while the remaining four states served for validation. The results indicate that (1) U-Net effectively extracted tree crowns from aerial imagery to construct the sample dataset; (2) among the tested algorithms, XGBoost achieved the highest accuracy of 0.88 when incorporating Sentinel-1, Sentinel-2, vegetation indices, and land cover features, outperforming models using only Sentinel-2 data by 25.7%; and (3) XGBoost-estimated tree canopy cover (Model TCC) showed finer spatial details than the National Land Cover Database Tree Canopy Cover (NLCD TCC), with R² against the true tree canopy closure from U-Net (True TCC) up to 49.1% higher. This approach offers a cost-effective solution for regional-scale canopy monitoring.