Land Surface Temperature （LST） is a key parameter in the study of global climate change. Thermal infrared remote sensing LST products are an ideal data source for obtaining LST. However， at present， LST products suffer from large-area deletions caused by clouds， and the limited time series cannot meet the needs of climate change research in historical periods， which limit the in-depth application of LST products. In this paper， the cumulative solar radiation SOA of ERA5 is used to characterize the change of LST under cloud， and three factors （radiative factor， terrain factor and spectral factor） used for LST reconstruction are combined to construct a random forest reconstruction model of Gap-Free LST， and the model is discussed. Reconstruction effects in cloudy sky， and time-series migration. The research results show that： （1） the factors used for LST reconstruction are ranked by importance， and the topographic and radiative factors show high importance.and the topographic and radiative factors show high importance. The LST reconstruction model constructed has a high degree of fit， R² is 0.97， RMSE is 1.27 K. （2） The reconstructed cloud-air Gap-Free LST fixes the fragmentation of the original LST distribution. Verified by ground station LST， R² is above 0.90. RMSE is between 2.67 K and 3.15 K. Comparing the change trend of SOA and ground station LST， it is found that the two show good coherence， indicating that SOA can fully reflect the change of LST under the cloud. （3） The seamless LST with sequential migration is reconstructed. For the LST with monthly sequential migration， R² ranges from 0.77~0.96 and RMSE ranges from 1.35 K~4.02 K. For LST with annual sequential migration， R² is above 0.86 and RMSE is between 2.73 K and 3.25 K. combined with the statistical map of predictor variables， it is found that the radiation factor is less affected by time migration， and the spectral factor NDVI migrates with time A change occurs， when the range of the NDVI used for LST reconstruction and the NDVI used for LST reconstruction model training is greatly different， the accuracy of the reconstructed LST will be reduced.This study can provide some theoretical support for long-time and seamless LST reconstruction.