Urban land-use prediction plays a pivotal role in sustainable development, infrastructure planning, and environmental monitoring. However, the task remains challenging due to the intricate features in high-resolution satellite imagery and the high visual similarity among urban land-use categories. To address these challenges, we propose LightSE-ResCNN-21, a novel lightweight convolutional neural network that combines depthwise separable convolutions, squeeze-and-excitation (SE) blocks, and residual connections to optimize both computational efficiency and classification performance. The model is trained on a custom dataset of 10,500 satellite images spanning 21 urban land-use classes, each resized to 256 × 256 pixels. To enhance generalization, we employ data augmentation, expanding the original 100 images per class to 500. LightSE-ResCNN-21 achieves a 96% classification accuracy, with an average precision of 90.5%, recall of 91.2%, and F1-score of 90.8%, while maintaining a compact size of ~5 MB and a fast inference time of ~11 ms per image. In comparative evaluations, our model significantly outperforms baseline architectures: VGG16 attains only 69% accuracy, while InceptionV3 reaches 90% under identical conditions. Detailed analysis of the confusion matrix and per-class metrics confirms the model’s robustness, particularly in discriminating visually similar categories such as golf courses, buildings, and storage tanks. Beyond standalone classification, we integrate CNN-based predictions with traditional machine learning algorithms and Cellular Automata (CA) to simulate spatial land-use transitions. This hybrid framework enables dynamic urban modeling with cloud-ready scalability, supporting real-time land-use forecasting. The system demonstrates strong potential for deployment in smart city planning and environmental monitoring, offering a balance of high accuracy, computational efficiency, and practical applicability.