UVSENet: A Semantic-Embedded dual branch fusion network for urban villages mapping

Urban villages (UVs) represent a typical yet highly complex spatial outcome of rapid urbanization. Accurate identification of UVs from remote sensing imagery is of great significance for understanding urban–rural mechanisms and supporting scientific urban governance and sustainable development. However, traditional optical remote sensing imagery is inherently limited by insufficient single modality representation capability, while multimodal data fusion is often constrained by data heterogeneity and cross modal semantic inconsistency. To address these challenges, this study proposes a cross modal semantic guided dual branch network for UVs identification (UVSENet). The proposed model deeply integrates Satellite Embedding data with Sentinel-2 optical imagery, enabling semantic level information compensation and structural level joint perception. Specifically, the semantic branch encodes high-level prior knowledge to guide the optical branch in achieving semantic alignment and feature enhancement during spatial structure extraction. This design further facilitates cross regional knowledge generalization, thereby improving the recognition of complex spatial patterns characteristic of UVs. Experimental results demonstrate that the proposed model consistently outperforms conventional single data source methods and shallow feature fusion approaches in terms of IoU, F1 score, and overall accuracy. The model achieves an IoU of 71.44 % and an F1 score of 83.35 % for UV identification, outperforming existing methods and demonstrating highly competitive performance. These findings validate the effectiveness of incorporating satellite semantic embedding data into remote sensing semantic segmentation and highlight its potential for large scale UV mapping. Moreover, this work provides a novel technical pathway for multimodal semantic alignment and knowledge driven urban analysis.