Reconstruction of MODIS Land Surface Temperature Products Based on Multi-Temporal Information

Land surface temperature (LST) products derived from the moderate resolution imaging spectroradiometer (MODIS) sensor are one of the most important data sources used to research land surface energy and water balance at regional and global scales. However, MODIS data are severely contaminated by cloud cover, which limits the applications of LST products. In this paper, based on the spatio-temporal autocorrelation of land surface variables, a reconstruction algorithm depending on the correlations between spatial pixels in multiple time phases from available MODIS LST data is developed to reconstruct clear-sky LST values for missing pixels. Considering the impacts of correlation and bias between predictors and reconstructed data on the modeling error, the known data in the reconstructed time phase are combined with the data temporally nearest to them as predictor variables to establish their temporal relationships with the reconstructed data. The reconstructed results are validated by a series of evaluation indices. The average correlation coefficient between the reconstructed results and ground-based observations is 0.87, showing high temporal change accuracy. The difference in Moran’s I, representing spatial structure characteristics between the known and reconstructed data, is 0.03 on average, indicating a slight loss of spatial accuracy. The average reconstruction rate is approximately 87.0%. The modeling error, as part of the reconstruction error, is only 1.40 K on average and accounts for 5.0% of the total error. If the product and modeling errors are removed, the residual error represents approximately 3.5 K and 5.6 K of the annual mean difference between the cloudy and cloudless LST at night and during the day, respectively. In addition, different reconstruction cases are demonstrated using various predictor data, including many combinations of multi-temporal MODIS LST data, the microwave brightness temperature, and the combination of the normalized difference vegetation index and terrain data. Comparisons among cases show that the known MODIS LST data are more reliable as predictor variables and that the data combination advocated in this paper is optimal