关键词: 地表温度, 热红外遥感, 降尺度, 机器学习, 随机森林, 解释变量, MODIS, 鲁棒性

Abstract:

Land Surface Temperature (LST) is one of the most important environmental parameters that describe the atmospheric and land surface interactions and energy balance in the terrestrial ecosystem at regional and global scales. Due to the trade off between high temporal and high spatial resolutions of remotely sensed LST products, many downscaling algorithms have been developed. However, the selection of explanatory variables of LST and downscaling models are often restricted by study location, which limits the generalization performance of these models. In this paper, a robust downscaling method of land surface temperature using Random Forest algorithm (RRF) is proposed after the evaluation of relationships between six variables, including surface reflectance, spectral indices, terrain factors, land cover types, longitude and latitude information, andatmospheric reanalysis data, to establish a nonlinear relationship between LSTs and other land surface parameters. This paper selects 11 regions of China as study areas, the Moderate Resolution Imaging Spectroradiometer (MODIS) LST product is downscaled by RRF from 1 km to 90 m. A comparison with two other downscaling methods (the Basic-RF model and the thermal sharpening (TsHARP) algorithm) is also made. Taking Beijing city as a presentative study area, the RRF model is proved to achieve a more satisfied performance in both study area A and B, in which the RMSEs are 2.39 K and 2.27 K, respectively. Besides, the Root Mean Squared Error (RMSE) of the RRF model trained ins tudy area B and evaluated in study area A is 2.56 K, while the RMSE trained in study area A and evaluated in study area B is 2.44 K, with a small decreasein RMSE (i.e., 0.17 K). Our further experiment results prove the robustness of RRF model trained in a specific region while being applied to other regions, and indicate that we can downscale LST with the RRF model in a large area using a few study areas for model training.

Key words: land surface temperature, thermal infrared remote sensing, downscaling, machine learning, random forest, predictor variable, MODIS, robustness