An Integrated Object- and Pixel-Based Residual Compensation Framework for Land Surface Temperature Downscaling

The land surface temperature (LST) downscaling is a valuable technique for obtaining high-spatiotemporal resolution LST data. By incorporating object-level or pixel-level residuals, the spectral information of the downscaled LST can be better restored, leading to improve the prediction accuracy. However, pixel-level residuals may introduce uncertainty in the residual results, whereas object-level residuals may overlook detailed land cover change information within objects, potentially resulting in discontinuities and abrupt changes in the predicted results. To address these challenges, this study proposed an integrated object- and pixel-based residual compensation framework (IOPRCF) to effectively restore the land cover change information within fine-resolution LST, thereby generating more accurate high-temporal and spatial resolution LST data. The IOPRCF was rigorously tested in 12 geographical diverse regions in China using three methods: thermal sharpening algorithm (TsHARP), random forest (RF), and simple and effective downscaling (SED). These tests encompass multiple time intervals and scaling scales to comprehensively validate the effectiveness and applicability of IOPRCF in the field of LST downscaling. The results demonstrate that, when compared with the original benchmark algorithms, TsHARP, RF, and SED with IOPRCF achieved average increases in <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> of 0.035, 0.035, and 0.028, respectively, average decreases in root mean square error (RMSE) of 0.110, 0.120, and 0.133 K, respectively, and average increases in SSIM of 0.017, 0.014, and 0.016, respectively. Moreover, the downscaled LST data produced by these algorithms with IOPRCF exhibited more similar spatial structures and captured more detailed information. Furthermore, the proposed IOPRCF demonstrated strong robustness and transferability across different LST downscaling algorithms and segmentation models. Finally, IOPRCF is expected to enhance the capabilities of LST downscaling algorithms for recovering land cover information, effectively supporting the generation of high-quality, high-spatiotemporal resolution LST globally.