Benchmarking deep learning architectures for land-cover semantic segmentation: insights from the DeepGlobe dataset

Accurate semantic segmentation of satellite imagery is essential for land-use analysis, environmental monitoring, and disaster management. This study systematically benchmarks five deep learning architectures—U-Net, PSPNet, DeepLabV3+, SegNet, and FCN-8s—using the DeepGlobe Land-Cover Classification dataset under unified experimental conditions. All models were evaluated using consistent preprocessing, optimisation, and training configurations. Performance assessment incorporated pixel accuracy, mean Intersection-over-Union (mIoU), Dice coefficient, precision, recall, class-wise analysis, confusion matrices, and computational efficiency. DeepLabV3+ achieved the highest segmentation performance, with an mIoU of 63.04%, Dice coefficient of 66.88%, and inference latency of 0.0788 s/image, demonstrating effective multi-scale feature extraction and boundary delineation. PSPNet and FCN-8s exhibited competitive segmentation performance, while U-Net and SegNet offered alternative accuracy–efficiency trade-offs. A resolution ablation study further demonstrated that increasing input resolution from 144 × 144 to 256 × 256 improved segmentation performance at a moderate computational cost. Additionally, loss-function and qualitative analyses provided insights into architectural behaviour across heterogeneous land-cover regions. Rather than introducing a new architecture, this study establishes a reproducible, multidimensional benchmarking framework and provides practical model-selection guidelines for remote sensing applications with varying accuracy and computational requirements, thereby supporting informed architecture selection for operational land-cover mapping tasks.