Land Surface Temperature Reconstruction Model of FY-4A Cloudy Pixels Considering Spatial and Temporal Characteristics

  Objectives  Resorting to remote sensing technology, land surface temperature (LST) data of high spatial and temporal resolution can be conveniently acquired. However, due to the influence of bad weather conditions such as cloud and rain, there are many null values in those LST data, being an obstacle to the further application of LST data.

  Methods  A novel null reconstruction model of LST based on spatial and temporal characteristics is proposed.Firstly, the reconstruction of partial null values of LST data is realized by weighting the effective records of the same period within contiguous days in the time domain, followed the intrinsic assumption that the changes of LST data in adjacent time periods are similar. Secondly, the reconstruction of other null valued pixels are reconstructed in the spatial domain, based on the assumption that adjacent similar pixels have similar LST. The similar pixels are searched by the normalized vegetation index (NDVI). In addition, due to the influence of extreme weather conditions, the above two steps are repeated several times until all null pixels are entirely reconstructed. Finally, Savitzky-Golay (S-G) filter is employed to remove the noise of the reconstructed LST data.

  Results  Compared to the in-situ data, the root mean squared errors (RMSE) of the LST data of all the FY-4A before and after reconstruction are 4.805 K and 6.969 K, respectively. Their determination coefficients (R²) are 0.842 and 0.605, respectively. Under clear sky conditions, RMSEs of FY-4A LST data before and after reconstruction are 3.392 K and 5.016 K, respectively, and their R² are 0.948 and 0.874, respectively. Under cloudy weather conditions, RMSE and R² of reconstructed LST data are 5.053 K and 0.726, respectively. Under rainy weather conditions, RMSE and R² of reconstructed LST data are 7.872 K and 0.313, respectively. Compared with the original LST data of FY-4A, the RMSE of the reconstructed data ranges 0.483-0.507 K and the R² ranges 0.846-0.976 for the null value regions of different sizes in spatial domain. RMSE ranges 0.405-1.915 K and R² ranges 0.952-0.989 for invalid LST time series of different sizes in time domain.

  Conclusions  The proposed null reconstruction model of LST can accomplish effective reconstruction not only for clear weather, but also for cloudy weather of long time series. Excluding the error of FY-4A LST data, RMSE of reconstruction results reaches 2.171 K. When the number of LST valid pixels is very small, the invalid LST can also be effectively reconstructed by the proposed model, which is tough for the diurnal temperature cycle (DTC) model.