High-Resolution Urban Surface Upward Longwave Radiation Estimation Using a Physics-Constrained Neural Network With Sky View Factor

Abstract

Surface upward longwave radiation (SULR) plays a critical role in characterizing urban thermal environments and quantifying urban heat island intensity. However, existing SULR retrieval methods are primarily designed for relatively homogeneous surfaces and often rely on coarse-resolution satellite products, limiting their ability to capture the radiative effects of complex urban morphology. To address these limitations, this study proposes a physics-constrained urban SULR estimation framework that integrates radiative-transfer theory, urban morphological information, and high-resolution remote sensing observations. Using Landsat 9 imagery, the framework combines physically meaningful variables, including land surface temperature (LST), emissivity, atmospheric radiation components, and sky view factor (SVF), within a unified learning architecture. Unlike purely data-driven approaches, the model incorporates radiative-transfer consistency as a physical constraint during optimization, improving both predictive accuracy and physical plausibility. In addition, SVF is introduced to explicitly characterize 3-D urban geometry and its influence on longwave radiation exchange. Experiments conducted across urban regions with different building densities demonstrate that the proposed framework effectively captures morphology-dependent radiation processes and achieves high retrieval accuracy, with errors as low as 4.6 W/m2 in dense urban environments. The results further reveal the important role of urban morphology in regulating longwave radiation and highlight the benefits of integrating physical constraints into urban thermal remote sensing. Overall, the proposed framework enables 30-m resolution urban SULR mapping and provides a practical approach for high-resolution urban energy balance and heat island studies.