MVNet: a building extraction method based on multiscale vision aggregation

Building extraction from high-resolution remote sensing (RS) images constitutes a fundamental task in geospatial applications such as urban planning, disaster response, and land use analysis. Recent deep learning-based encoder-decoder architectures have significantly improved extraction performance. Nevertheless, prevailing approaches still struggle with complex urban scenes where buildings exhibit substantial variations in scale, shape, and spatial distribution. To address these challenges, we propose MVNet, a novel multi-branch network that simultaneously augments contextual comprehension and structural characterization of buildings. MVNet employs an encoder-decoder architecture and introduces two key modules: the Contextual Feature Aggregation (CFA) module dynamically fuses global and local semantic features to enhance foreground-background discrimination, while MSA sharpens the delineation of multi-scale features and boundary details through attention mechanisms. These modules are seamlessly integrated to reinforce feature representations from both semantic and structural perspectives. Extensive experiments on three publicly available benchmarks, including WHU, Massachusetts, and Inria, demonstrate that MVNet surpasses contemporary state-of-the-art approaches by margins of 1.2-3.5% in IoU. Quantitative and qualitative analyses jointly substantiate the effectiveness of the proposed modules in handling diverse urban environments and complex building distributions.