Correct land-cover classification is important for environmental monitoring, urban planning, and disaster management. Conventional classification techniques tend to fail in the complexity of high-resolution aerial imagery. In this research, we present a method based on deep learning using the U-Net architecture for semantic segmentation of land-cover classes from the DeepGlobe Land-Cover Classification Dataset. The U-Net model has an encoder-decoder structure with skip connections. It effectively captures both local and global context information, making it suitable for the pixel-wise classification task. We preprocess the dataset by resizing images, normalizing pixel values, and converting segmentation masks into categorical classes. We train the model using the categorical cross-entropy loss function and optimize it with the Adam optimizer. Experimental results show that our method achieves high classification accuracy across various land-cover classes and outperforms traditional machine learning techniques. These results indicate that deep learning methods, especially U-Net, can be strong tools for remote sensing applications.