Comparative analysis of deep learning models for semantic segmentation and classification on high-resolution aerial imagery

High-resolution aerial imagery plays a vital role in areas like urban planning, environmental monitoring, and guiding autonomous systems. However, achieving precise segmentation of these images is challenging due to variations in object sizes, uneven class distribution, and intricate boundaries. In this work, we evaluate the real-time segmentation performance of four distinct deep learning architectures such as U-Net, DeepLabV3+, ResNet-50, and EfficientNet-B4 on the Semantic Drone Dataset. To enhance robustness and training stability, we applied Albumentations-based data augmentation along with the OneCycleLR learning rate scheduler. Performance was measured using metrics including mean Intersection over Union (mIoU), precision, recall, and processing speed (FPS). Among the tested models, EfficientNet-B4 produced the highest scores, attaining an mIoU value of 0.6123 and a speed of 705.48 FPS, making it well-suited for real-time drone applications. ResNet-50 provided a strong accuracy–efficiency balance, while U-Net and DeepLabV3+ presented trade-offs between detail preservation and processing speed. Overall, this study offers a practical performance comparison that addresses both precision and real-time viability for aerial segmentation in real-world scenarios.