"data_raw_test"

"For a better grasp of the ecosystem services cities offer and to support environmental sustainability, precisely segmenting Urban Blue-Green Spaces (UBGS) is crucial. Characterizing the complexity of urban spatial morphology, coupled with the different distribution of buildings, water and vegetation, poses significant challenges for segmenting UBGS from high-resolution imagery. To overcome the continuous limitations of present methods in dealing with blurred landscape boundaries and spectral feature confusion, this study proposes Multi-scale Dualbranch Attention Network (MDANet). MDANet constructed a multi-scale dual-branch attention network, which couples feature pyramids with a channel attention mechanism for cross-scale spatial-spectral feature. On the other hand, an enhanced multimodal dataset, E-BGfield, is constructed by integrating Sentinel-2 imagery with NDVI and MNDWI for model performance. Experimental results demonstrate the performance of MDANet with 74.6% MIoU and 90.6% mAP@50 on E-BGfield. Spatialtemporal analysis conducted in Nanjing, an important city in the Yangtze River Basin and the capital of Jiangsu Province, from 2015 to 2024 reveals a policy intervention-landscape response coupling effect. The increase in plant area is 0.58 km2, while water area recovered to 1.82 km2 after ecological restoration following initial shrinkage. Change Intensity Index (CII) indicates a 12\u201318 months lag between policy implementation and ecological response. The technical framework proposed by this research provides innovative methods and extensive data support for ecological governance and decision making in sustainable heterogeneous urban development.  "