LGI-UNet: a robust semantic segmentation framework with local–global interaction for urban green space mapping in mountainous cities

Urban green spaces (UGSs) must be accurately extracted to support refined urban management and sustainable development. However, UGSs in mountainous environments are scattered and exhibit multi-scale heterogeneity, hindering current deep learning-based segmentation methods. In response, this study constructed a novel semantic segmentation framework, the Local–Global Interaction Network (LGI-UNet), which designs a Difference-Aware Attention (DAA) module to model the differences between local features and global statistical context, thereby enhancing the discriminative representation of shaded and spectrally ambiguous regions. A high-resolution Unmanned Aerial Vehicle dataset covering various mountainous urban scenarios was constructed, and comparative experiments were conducted with representative models based on convolutional neural network, Transformer, and Mamba. LGI-UNet demonstrated superior performance on multiple evaluation metrics. Paired t-test results confirm that LGI-UNet significantly outperforms the best baseline model and effectively identifies green spaces in complex scenarios. The model showed consistent performance in spatially unsampled areas within the same urban environment. LGI-UNet provides an effective and robust solution for UGS extraction in complex mountainous environments, offering valuable technical support for ecological monitoring and urban planning.