Downscale MODIS Land Surface Temperature Based on Three Different Models to Analyze Surface Urban Heat Island: A Case Study of Hangzhou

Remote sensing technology plays an increasingly important role in land surface temperature (LST) research. However, various remote sensing data have spatial–temporal scales contradictions. In order to address this problem in LST research, the current study downscaled LST based on three different models (multiple linear regression (MLR), thermal sharpen (TsHARP) and random forest (RF)) from 1 km to 100 m to analyze surface urban heat island (SUHI) in daytime (10:30 a.m.) and nighttime (10:30 p.m.) of four seasons, based on Moderate Resolution Imaging Spectroradiometer (MODIS)/LST products and Landsat 8 Operational Land Imager (OLI). This research used an area (25 × 25 km) of Hangzhou with high spatial heterogeneity as the study area. R2 and RMSE were introduced to evaluate the conversion accuracy. Finally, we compared with similarly retrieved LST to verify the feasibility of the method. The results indicated the following. (1) The RF model was the most suitable to downscale MODIS/LST. The MLR model and the TsHARP model were not applicable for downscaling studies in highly heterogeneous regions. (2) From the time dimension, the prediction precision in summer and winter was clearly higher than that in spring and autumn, and that at night was generally higher than during the day. (3) The SUHI range at night was smaller than that during the day, and was mainly concentrated in the urban center. The SUHI of the research region was strongest in autumn and weakest in winter. (4) The validation results of the error distribution histogram indicated that the MODIS/LST downscaling method based on the RF model is feasible in highly heterogeneous regions.