Spatial Downscaling Study of Land Surface Temperature Based on Multilayer Perceptron Model

摘要

Land Surface Temperature (LST) plays an important role in characterizing surface energy conversion and climate.Currently,there is a contradiction between temporal resolution and spatial resolution of commonly used LST data sources.With Xi’an City as the research object,Multilayer Perceptron (MLP) model was used to downscale 1,000 m × 1,000 m LST product to 250 m × 250 m based on MODIS data.The fitting effect was compared with that of traditional multiple linear regression model,and the LST retrieved from Landsat 8 OLI_TIRS was used as the reference to evaluate the accuracy.The results showed that the R~2 of LST data fitted by MLP model in the daytime and at night was above 0.85,and the predicted residuals followed the normal distribution.The model had good fitting effect,and the fitting effect of LST in the daytime was better than that at night,while the output LST was lower than the original LST.Compared with multiple linear regression model,the R~2 of MLP model was larger and the RMSE was smaller both in the daytime and at night.The MLP model had not only stronger explanatory ability,but also more accurate prediction results.After downscaling by MLP model,the spatial resolution of LST was improved,which could reflect the spatial distribution pattern of LST and landscape features of underlying surface more accurately.The test results of LST retrieved from Landsat 8 OLI_TIRS showed that the covariance of the two was positive and the correlation coefficient was 0.951 3.The MLP model achieves the expected downscaling effect well,and has important application significance in acquiring high-resolution LST.