A flexible framework for identifying urban villages using Sentinel-2 observations and deep learning

Abstract

Identifying urban villages (UVs) is essential for supporting urban renewal and governance, as these informal areas present significant challenges to urban sustainability. Existing multi-modal deep-learning studies have accurately identified UVs in individual cities; however, their scalability is constrained by substantial computational demands and data availability. Currently, efficient deep-learning models tailored for large-scale, consistent UV maps using freely released Sentinel-2 data remain underexplored. Therefore, we proposed a flexible framework leveraging multiple Sentinel-2 observations and deep-learning models to cost-effectively support large-scale UV mapping. We specifically employed a dual-branch encoder to consolidate discriminative features: one branch extracted static information from annual images, while the other captured seasonal dynamics induced by building shadows across four seasonal images. These features were then jointly represented using fusion modules and used for UV identification through skip connections. The model was trained and tested on self-compiled Guangzhou-Shenzhen samples, with the Pearl River Delta serving to analyze applicability and transferability. Results demonstrated incorporating dynamic information notably enhanced the accuracy of UV identification. The proposed framework exhibited optimal performance (mIoU: 83.33 %) with robust spatial generalization, surpassing traditional methods (mIoU improvements: 0.46–12.49 %). Compared to multi-modal data-driven studies, our framework can be conveniently implemented in complex urban environments while maintaining both efficiency and accuracy in large-scale applications. This cost-effective solution offers high-quality, large-scale UV mapping, thereby advancing UV management and planning efforts.