QE-WNet: A Quadratic Encoding-decoding W-shape Network for Road Segmentation based from Optical Remote Sensing Imagery

With advancements in imaging technology, high-resolution remote sensing images are now readily available for applications such as urban planning, resource assessment, and land cover analysis. Roads are a crucial element of infrastructure, and extracting road information from high-resolution remote sensing images is vital for tasks like urban planning and mapping. However, existing road extraction techniques, which rely on traditional encoder-decoder architectures, often struggle to accurately delineate road boundaries and surfaces. As a result, these methods tend to produce subpar outcomes and are challenging to implement in practical systems. To address this issue, we introduce a novel quadratic encoding-decoding W-shaped network, referred to as QE-WNet. This approach recognizes the differing characteristics of road bodies and boundaries—where road boundaries are more stable in semantic representation and road bodies are less so. In the first stage of encoding and decoding, we focus on modeling road boundaries, while the second stage emphasizes road bodies, leveraging the more stable secondary encoding-decoding information. This separation enhances the accuracy of road extraction. We conduct comparative experiments using several established methods on the Deepglobe dataset, and the results demonstrate the effectiveness of our proposed method.