Urban open space extraction with large kernel dual-path convolutional network from high-resolution remote sensing images

• Published in: International journal of remote sensing Vol. 46; no. 20; pp. 7707 - 7728
• Main Authors: Zhou, Pengyao, Li, Yanfei, Li, Xianju
• Format: Journal Article
• Language: English
• Published: London Taylor & Francis 18.10.2025; Taylor & Francis Ltd
• Subjects: Cities; Datasets; Deformation; dual-path convolutional network; Formability; Geometric transformation; High resolution; high-resolution remote sensing images; Image resolution; Modelling; Modules; Open spaces; Remote sensing; Semantic segmentation; Urban areas; Urban environments; Urban open space; Urban planning
• ISSN: 0143-1161; 1366-5901
• DOI: 10.1080/01431161.2025.2559420

Urban open spaces refer to public outdoor spaces within cities that are vital for urban planning and the environment. These spaces feature complex landscapes with diverse ground objects, and the varying spatial scales of different types of open spaces pose challenges for conventional convolutional models or attention mechanisms, limiting their precision. This paper introduces a new semantic segmentation network, PYNet, designed for high-precision extraction of urban open spaces from high-resolution remote sensing images. PYNet incorporates a lightweight Large Selective Kernel Network (LSKNet) as the backbone, which dynamically adjusts the receptive field to efficiently handle complex spatial contextual information. Additionally, an innovative Deformable Dual-Path Attention Module (DDPAM) combines a newly designed Transformation Specialization Module (TSM) and a lightweight attention mechanism to enable fine-grained modelling of both global and local features in parallel. The TSM uses deformable convolutions to model geometric transformations, which enhances the delineation of irregular object boundaries, achieving high-precision extraction of urban open spaces. PYNet achieved 87.24% mIoU on the ISPRS Potsdam benchmark dataset and 93.00% mIoU on the experimental dataset covering 20 cities, outperforming the second-best model by 0.74% and 1.3%, respectively. Experimental results demonstrate that PYNet significantly improves extraction accuracy by overcoming the limitations of traditional attention mechanisms in transformation modelling capacity, showing robustness to diverse urban landscapes.