Land surface temperature (LST) is a key variable at the land–atmosphere boundary. For many research projects and applications an all-weather LST product at moderate spatial resolution (e.g., 1 km) would be highly useful, especially in frequently cloudy areas. Merging thermal infrared (TIR) and microwave (MW) observations is able to overcome shortcomings of single-source remote sensing to derive such an LST. However, in current merging methods, models adopted for downscaling MW LST fail to quantify the effect of temporal variation of LST. Thus, accuracy of the merged LST can be deteriorated and therefore remain a major impediment for these methods to be generalized over large areas. In this context, we propose a new practical method to merge TIR and MW observations from a perspective of decomposition of LST in temporal dimension. The physical basis of the method is decomposing LST into three temporal components: annual temperature cycle component, diurnal temperature cycle component prescribed by solar geometry, and weather temperature component driven by weather change. The method was applied to MODIS and AMSR-E/AMSR2 data to generate an 11-year record of 1-km all-weather LST over Northeast China: the resulting merged LST has an accuracy of 1.29–1.71 K when validated against in situ LST; besides, no obvious differences in accuracy of the merged LST were found between clear-sky and unclear-sky conditions. Furthermore, the proposed method outperforms the previous method in both accuracy and image quality, indicating its good capability to generate daily 1-km all-weather LST, which will benefit continuous monitoring of earth’s surface temperature.