Year: 2025
Type: conference-paper
Abstract: To address current limitations in semantic segmentation models for building extraction from high-resolution remote sensing imagery—including structural voids in large-scale buildings, suboptimal edge segmentation, and inadequate detection of small structures—we propose MEU-Net, an encoder-decoder architecture enhanced through three strategic innovations. The framework incorporates: (1) a Multi-scale Feature Enhancement Module (MFEM) to improve extraction accuracy across varying building scales through hierarchical receptive field fusion; (2) a Feature Fusion Module (FFM) designed to refine edge delineation and ensure structural completeness of large buildings via cross-level feature integration; and (3) a modified loss function incorporating class-balancing mechanisms to mitigate negative impacts from imbalanced pixel distribution. Experimental evaluations on the WHU aerial image datasets and Massachusetts building datasets demonstrate that MEU-Net achieves significant improvements over the baseline U-Net, with IoU metrics increasing by 3.61% and 2.51% respectively. These results validate the model's capability to enhance extraction accuracy for buildings of varying scales while maintaining robust edge delineation and spatial coherence, offering a promising solution for urban planning and land-use analysis applications.
Authors: Songyi Liu, Yingze Fang, Zhe Geng, Ziwei Lv, Caitang Sun
Institution: Jilin University
Language: English
Subfield: Ocean Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413845248