With the ever-growing expansion and densification of urban landscapes, accurate land use classification and tracking have become indispensable tools for effective and sustainable urban planning. This study tackles the challenge of sustainable urban development by employing three different deep learning models: the Edge-Attention U-Net, ResUNet, and UNet-FS. These models enable pixel-wise classification, facilitating precise allocation of various land use categories, including buildings, roads, vegetation, water bodies, and other crucial urban features in densely populated areas, exemplified by Dubai City. A comprehensive evaluation framework, encompassing Intersection over Union (IoU), Dice coefficient, Pixel accuracy, and class-specific metrics, is employed to assess the performance of the models. The comparative analysis reveals that the UNet-FS model outperforms the others, achieving a remarkable Jaccard coefficient of 0.9303, a Dice coefficient of 0.9637, and a pixel accuracy of 0.9885. The proposed Edge-Attention U-Net approach demonstrates exceptional accuracy in segmenting complex urban scenes and generating high-resolution land use maps for urban planning purposes. The incorporation of attention mechanisms enhances feature localization and reduces noise, contributing to reliable segmentation outcomes. This research advances remote sensing and urban studies by providing an efficient solution for land use classification and facilitating informed urban development strategies. The UNet-FS model’s exceptional performance in land use classification supports decision-making in urban development and sustainability endeavors. In conclusion, this study presents an innovative urban planning and monitoring framework by leveraging deep learning and attention-based semantic segmentation.