A novel dual-layer composite framework for downscaling urban land surface temperature coupled with spatial autocorrelation and spatial heterogeneity

Land surface temperature (LST) captures fundamental information on the spatiotemporal variation of energy balance at the surface. The trade-off between spatial and temporal resolutions of remote sensing images (retrieved LSTs), however, restricts fine-scale thermal environmental investigations. In this context, a novel dual-layer composite framework (DCF) for LST downscaling coupling spatial autocorrelation and spatial heterogeneity was developed based on the two fundamental laws of geography and used to improve existing kernel-driven methods. Besides, a new non-parametric kernel-driven LST downscaling method (N-DLST) was also proposed under the DCF, in which Bayesian non-parametric general regression (BNGR) was applied to predict the high-resolution LSTs with auto-selected kernels. In the experiment of downscaling Landsat 8 LST from 300 m to 30 m over the highly heterogeneous urban area, the N-DLST method significantly outperformed the original kernel-driven methods, with the highest coefficient of determination (R2 = 0.93) and lowest root mean square error (RMSE = 0.85). Moreover, the enhanced effects of DCF in downscaling LST were demonstrated by comparing the accuracy of the disaggregation of radiometric surface temperature (DisTrad), the geographically weighted regression-based method (GWR), and random forest (RF) method before and after their improvements. Visual interpretation and quantitative assessments revealed that the DCF could improve the accuracy of DisTrad, GWR, and RF methods with an increase in R2 by approximately 0.09 and a decrease in RMSE by more than 0.4 °C. In the cases of LST downscaling over highly heterogeneous contexts and water bodies, N-DLST effectively preserved the textures and large-scale variations, yielding the most consistent spatial pattern with the reference LST. Given the simplicity of the modelling process and absence of auxiliary data, the DCF could strengthen the performance of both linear and nonlinear LST downscaling methods, while the N-DLST method could serve as an effective tool for high-resolution LST prediction.