Accurately estimating crop cultivation areas is critical for predicting yields and managing overproduction, particularly for staple crops grown in regions like Jeju Island, South Korea, where reporting cultivation areas is mandatory. This study developed a modified U-Net architecture for semantic segmentation, utilizing UAV-based high-resolution imagery in the open-source NIA AI HUB dataset. The dataset includes labeled RGB images of six winter crops—white radish, cabbage, onion, garlic, broccoli, and carrot—grown on Jeju Island, a key agricultural hub. The proposed model incorporates a ResNet-34 backbone, Attention Gates, and Residual Modules, achieving a mean F1 score of 85.4% and an intersection over union (IoU) of 74.6%, outperforming the original U-Net. This advancement significantly reduces misclassifications among visually similar crops, such as garlic and onion. Application to three unknown fields demonstrated a mean prediction accuracy of 90.2%, effectively estimating cultivation areas with high precision. By leveraging public datasets and innovative AI techniques, this study highlights the scalability and practicality of the proposed model in enhancing precision agriculture. These findings demonstrate the model’s potential to improve crop yield prediction, optimize resource allocation, and support sustainable farming practices in diverse agricultural environments.