Accurate and timely crop mapping is essential for global food security assessments; however, conventional crop mapping models are usually applicable to specific spatial or temporal scales, i.e. “one-time, one-place” model. Moreover, the extensive application of a trained model to other regions is challenging when sufficient ground-truth samples used for training process are unavailable. This study exploited Cropland Data Layer (CDL) and Landsat data for Arkansas, United States (US), to train a U-Net model and then extensively tested the generalization ability of the model in the Corn Belt and California in the US, and even further tested in Liaoning, China, on a transcontinental scale. Two temporal images were generated by compositing the median values of images obtained during two crop growing time windows representing the sowing (from March to May) and vigorous growth periods (from June to August). In order to ensure the consistency of the data distribution between the target areas (testing areas) and the training area, we shifted the time windows of the target areas to match that of the training area following the phenological matching principle. Then we can composite the target data (testing data) according to the time windows matched in the target areas. The results showed a satisfactory accuracy. The average optimal Overall Accuracy of crop mapping in all the target areas exceeded 87%. The average optimal F1-score of corn and rice was 0.79. Finally, we compared the generalization performance of U-Net and Random Forest (RF) classifiers. The results showed that U-Net performed better in all the target areas in the US while RF performed better in target areas where the plots were smaller. The procedure and strategy developed will facilitate the realization of high-performance and automated global model transfer.