A gradient-based nonlinear multi-pixel physical method for simultaneously separating component temperature and emissivity from nonisothermal mixed pixels with DART

Component temperature and emissivity are crucial for understanding plant physiology and urban thermal dynamics. However, existing thermal infrared unmixing methods face challenges in simultaneous retrieval and multi-component analysis. We propose Thermal Remote sensing Unmixing for Subpixel Temperature and emissivity with the Discrete Anisotropic Radiative Transfer model (TRUST-DART), a gradient-based multi-pixel physical method that simultaneously separates component temperature and emissivity from non-isothermal mixed pixels over urban areas. TRUST-DART utilizes the DART model and requires inputs including at-surface radiance imagery, downwelling sky irradiance, a 3D mock-up with component classification, and standard DART parameters ( e.g. , spatial resolution and skylight ratio). This method produces maps of component emissivity and temperature. The accuracy of TRUST-DART is evaluated using both vegetation and urban scenes, employing Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images and DART-simulated pseudo-ASTER images. Results show a residual radiance error is approximately 0.05 W/(m 2 ·sr). In absence of the co-registration and sensor noise errors, the median residual error of emissivity is approximately 0.02, and the median residual error of temperature is within 1 K. This novel approach significantly advances our ability to analyze thermal properties of urban areas, offering potential breakthroughs in urban environmental monitoring and planning. The source code of TRUST-DART is distributed together with DART ( https://dart.omp.eu ). • Physically unmixing for simultaneous component emissivity-temperature separation. • The median residual radiance error is around 0.05 W/(m 2 ·sr). • Ideal experimental conditions: Median errors: emissivity ≈ 2 %, temperature ≈ 1 K. • Incorporating physics, geostatistics, graphics, and gradient-based learning.