Building segmentation from remote sensing imagery based on deep feature extraction and regularization optimization
Abstract
To address segmentation accuracy degradation caused by scale variations and complex backgrounds in remote sensing building extraction, this study systematically compares and optimizes four representative deep learning architectures— FCN, U-Net, SegNet, and DeepLabV3—on the AI Earth Land Cover Extraction Dataset under a unified experimental framework. Results show that DeepLabV3 achieves the best baseline performance (IoU = 0.790) owing to its Atrous Spatial Pyramid Pooling module, followed by FCN (0.780), U-Net (0.765), and SegNet (0.720). To improve SegNet’s limited performance, grid-based hyperparameter optimization is conducted: replacing the VGG-16 backbone with ResNet- 101 and refining Dropout (0.3) and learning rate (1×10⁻⁴). The optimized SegNet achieves the highest IoU of 0.824 among all models. These findings demonstrate that strong feature representation combined with appropriate regularization critically enhances building segmentation accuracy, providing practical guidance for semantic segmentation model design in remote sensing applications.