Land-Unet: A deep learning network for precise segmentation and identification of non-structured land use types in rural areas for green urban space analysis

Abstract

Land Use and Land Cover Change (LUCC) have become popular research topics in the environmental field. With the development of artificial intelligence technology, many downstream applications based on intelligent urban–rural semantic analysis have emerged. Scholars have made significant progress in the intelligent analysis of urban imagery, but exploration of unstructured rural remote sensing data has been limited. This paper addresses the existing pixel-level semantic ambiguity issues and proposes a new deep learning model, Land-Unet. The network features a dual-branch Edge-Sensing Block (ESB) structure, including a Spatial and Channel Synergistic Attention (SCSA) branch and a Dynamic Upsampling (DYU) technique, which effectively resolves contour ambiguity in edge semantic information in rural images. Experiments on multiple datasets using various deep learning methods show that compared with the original structure, the proposed method increases $mIoU$ by 9.7%, $mDice$ by 5.9%, and $mAcc$ by 12.2%. Compared to transformer-based methods, proposed method also demonstrated improved performance. Additionally, a new rural satellite imagery dataset, RuralUse, has been open-sourced for semantic segmentation research.