Land surface temperature downscaling in urban area:A case study of Beijing

Land Surface Temperature (LST) is one of important variables for urban thermal environment studies. Urban surface is extremely complex and LST is heterogenous. High spatial resolution of LST is helpful for fine urban thermal environment monitoring and mitigation. However, so far, as we know, mostly LST downscaling studies focus on two-dimensional scope, and lack of building three-dimensional (3D) structure impact on LST. This study will use random forest model (RF) with both 2D and 3D land surface indices for downscaling of MODISF 1 km LST to 100m. Meanwhile, the spatial scale issues of building 3D morphology on LST is also discussed. In addition, in order to make up for the lack of theoretical basis of RF model, this study added more parameters during RF downscaling model generation based on the thermal radiation transmission equation, e.g. land surface radiance (MOD02) and precipitable water vapor (PWV, MOD05).The results show: (1) When MOD02 and MOD05 are included in RF model, RMSE and R2 between simulated 1 km LST and MODIS LST product are improved from 3.1 K and 0.5 to 0.38 K and 0.94. (2) When building 3D morphology is included in RF model, the OOB_score improves from 0.46 to 0.49. The R2 between simulated 100m LST and ASTER LST product is slightly decreased, one of the reasons is that LST retrieval methods of MODIS and ASTER are different and the two sensors are also different. However, when MOD02 and MOD05 are included, RMSE and R2 improve from 2.4 K and 0.29 to 1.2 K and 0.68. (3) The OOB_scores with building morphologies improve at both 1 km and 100 m scale, and the importance of building morphologies are different. Above all, downscaling MODIS LST in urban area should consider land surface radiance, PWV and building 3D structure indices, and impact of building morphologies on LST are different at different spatial scale. © 2021, Science Press. All right reserved.