DE-Net: A Density-Aware and Edge-Enhanced Network for High-Resolution Building Segmentation

Abstract High-resolution building extraction from remote sensing images plays a pivotal role in urban planning, disaster response, and land-use monitoring. However, the complex urban environment—characterized by dense building distributions, varied architectural styles, and blurred object boundaries—poses significant challenges for existing semantic segmentation models. To address these issues, we propose DE-Net, a novel semantic segmentation framework designed for precise building extraction from high-resolution unmanned aerial vehicle (UAV) imagery. DE-Net consists of three key components: a ConvNeXt-based feature backbone that captures hierarchical semantic features while preserving spatial detail; a Spatial Resolution Adaptive Reconstruction Module (SRARM) that dynamically decodes features based on predicted density maps, enabling tailored upsampling strategies for dense and sparse regions; and a Multi-Scale Edge-Aware Fusion Module (MS-EAM) that extracts edge attention from each encoder stage and fuses them to enhance boundary localization. Experiments on a high-resolution UAV dataset show that DE-Net outperforms other methods in IoU (86.78%), F1 score (92.91%), and boundary IoU, surpassing the second-best model, UPerNet, by 4.47% and 2.65%, achieving state-of-the-art performance.