Satellite-derived land surface temperature (LST) is a directional variable and has significant angular anisotropy. This characteristic contributes to enhance the differences among different satellite-derived LST products, and therefore increases the challenge of using multi-sensor and multi-decadal data to provide a long-term and angleconsistent LST climate data record. The kernel-driven model can balance the interpretability and operability well, so that it is suitable for angular normalization of LST products. The calibration of the kernel-driven model depends on multi-angle data which is difficult to obtain due to the spatial-temporal heterogeneity of LST. In this study, a novel LST angular normalization method based on the kernel-driven model was proposed to correct the angular effect of satellite-derived LST product by constructing multi-angle LST dataset from one geostationary satellite (GOES-R/ABI) and four polar-orbiting satellites (Terra/MODIS, Aqua/MODIS, Metop/AVHRR, and SNPP/VIIRS). The dataset gathered more abundant angle information, i.e., LSTs from three different observation geometries for the same pixel. The kernel-driven model was calibrated using the multi-angle LST dataset in the Continental United States (CONUS) during the year 2020. The discrepancies of the root mean square difference between LST before and after angular normalization range from 0.14 K to 1.10 K over nine land cover types in the four seasons. Similar results are obtained when the calibrated kernel-driven model was further expanded to other years and areas (i.e., the CONUS in 2021 and East Asia in 2020). The LST angular normalization method was applied to correct the angular effect of MODIS LST product. The results indicate that there is a strong correlation between the spatial distribution of LST differences (LST before and after angular normalization) and view zenith angle (VZA). MODIS LSTs before and after angular normalization were compared with Landsat 8 LST and Sentinel-3 A LST in near-nadir viewing for January, April, July, and October 2020. The angular normalization reduced the root mean square error (RMSE) between MODIS LST and Landsat 8 LST by 0.94-2.06 K in different months and by 0.13-2.61 K over various land cover types. For Sentinel-3 A, the RMSE decreased by 0.30-0.64 K in different months. The accuracies of MODIS LST before and after angular normalization were further validated using in situ measurements at the six SURFRAD sites. There are large discrepancies between the RMSE of MODIS LST before and after angular normalization versus in situ LST for VZA >= 45 degrees. The largest discrepancy is up to approximately 1.3 K at the GWN site. The LST angular normalization method has the potential to provide an angle-consistent LST climate data record.

来源平台：REMOTE SENSING OF ENVIRONMENT