For environmental monitoring, urban planning, and resource management, it’s important to be able to accurately classify land cover from satellite images. When it comes to spatial and contextual awareness, traditional pixel-wise categorization approaches typically have trouble, especially in landscapes that are diverse and varied. In this study, it suggested a new deep learning framework that combines U-Net architecture with improved attention mechanisms to classify land cover types in high-resolution satellite pictures. The U-Net’s encoder-decoder structure gathers contextual data at different scales, and channel and spatial attention modules are included into skip links to dynamically improve the relevance of features. This synergy makes it easier to tell the difference between classes and draw boundaries, particularly in areas with mixed or unclear land cover types. This can use benchmark satellite datasets to train and test our model, and it works better than baseline U-Net and other cutting-edge approaches. From the results obtained the proposed model in ISPRS Potsdam gave overall accuracy of 91.4%, mIoU of 78.5% and Mean F1 Score of 82.3% and in DeepGlobe gave overall accuracy of 89.1%, mIoU of 74.2% and Mean F1 Score of 79.5% respectively. The suggested model shows better generalization, noise resistance, and efficiency while dealing with images of different resolutions. Our method offers a solution for automated land cover mapping in remote sensing applications that can be scaled up and works well.