Urban land surface temperature prediction by integrating LSTM and geospatial information: A case study of Kunming, China

Abstract

The land surface temperature (LST) is a critical parameter reflecting the urban thermal environment. Accurate prediction of land surface temperature is essential for urban planning, heat island mitigation, and sustainable development. Existing models often do not capture both long-term temporal dependencies and spatial heterogeneity. This study proposes a feature-initialized spatial attention LSTM (FSA-LSTM) that integrates geospatial information to address these limitations. Key innovations include: (1) a spatial weight matrix to adaptively regulate hidden states, capturing local spatial dependencies; (2) geospatial feature-based initialization of hidden and cell states, enhancing convergence and stability; and (3) a spatial cross-attention mechanism that fuses hidden states with location information, enabling explicit interaction between model states and spatial context. When applied to old and new urban districts in Kunming, China, the FSA-LSTM outperforms baseline models, achieving $R^2$ improvements of 13.8–38.3% and 2.9–15.5%, respectively. Further evaluations, including cross-temporal predictions, diverse geographical scenarios (e.g., varying dominant land cover types, high altitudes, and sparse vegetation), and cross-regional experiments across 34 representative regions nationwide, indicate that FSA-LSTM exhibits strong transferability, robustness, and generalizability. Overall, the proposed FSA-LSTM provides a mechanistically informed, accurate, and scalable tool for urban thermal environment monitoring and management.