High-resolution local climate zone mapping via deep mixed-scene decomposition of remote sensing imagery

• Formulates Local Climate Zone (LCZ) mapping as a mixed-pixel unmixing task. • Proposes MSU-Net, a deep learning model for LCZ unmixing in complex urban areas. • Introduces a Dual Cross-Attention Fusion module to fuse multi-branch information. Under rapid urbanization, traditional single-class LCZ mapping methods fail to represent the coexistence of multiple land-cover types within the same block, resulting in blurred boundaries and reduced accuracy for urban heat-island modeling. To address this, LCZ mapping is reformulated as a mixed-scene unmixing task and tackled with a novel deep-learning framework, MSU-Net. Real street-block morphologies from OpenStreetMap are combined with 1 m Google Earth imagery to create multi-scale inputs that preserve both global block layouts and local detail. MSU-Net comprises a primary unmixing branch reinforced by two auxiliary guidance branches—one driven by purified-image semantic cues, the other by sparse local spatial reconstruction, and a Dual Cross-Attention Fusion (DCAF) module that integrates global–local and global–purified features under non-negativity and sum-to-one constraints. Two block-level datasets covering Wuhan and Shenzhen were created. MSU-Net outperforms existing methods on these datasets, boosting overall accuracy by 15–17 %, reducing mean absolute error by over 35 %, and cutting weighted-difference error by around 30 %. Transfer learning further confirms its robustness across cities with distinct morphologies.