Answers are grounded in this paper and cite their sources.
Despite the progress of deep learning in road extraction from VHR imagery, existing approaches often struggle with restricted receptive fields and insufficient global context modeling. These limitations become acute in complex urban environments, where shadows, occlusions, and inter-class similarities lead to road discontinuities and localization errors. These limitations hinder their ability to handle complex urban scenarios characterized by large road-scale variations, high visual similarity between roads and surrounding land-cover classes, and severe occlusions caused by shadows, vegetation, and high-rise buildings, often resulting in fragmented and inaccurately localized road networks. We propose SDFFNet, a feature fusion network that combines a ResNet encoder with a strip convolution decoder to solve these challenges. It integrates a large kernel cross-dimensional attention submodule to reduce shadow occlusion effects and enhance spatial context modeling ability. A fourier transform unit submodule captures global frequency domain information, while a spatial-frequency domain feature fusion module strengthens encoder-decoder interactions to mitigate interclass similarity interference. Additionally, a spatial pixel fusion module aligns deep semantic information to minimize prediction errors caused by pixel positional deviations. Experimental results show that SDFFNet achieves Intersection over Union (IoU) scores of 67.98%, 60.69%, and 64.85% on the Massachusetts, CHN6-CUG, and DeepGlobe datasets, respectively, outperforming comparable methods. The code for SDFFNet will be made available athttps://github.com/xuedanni1225/SDFFNet.
No takes yet. Share an insight, caveat, or question.
Xue et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: