Abstract Near‐space temperature is crucial for atmospheric dynamics and aerospace applications, yet data sets with refined vertical resolution remain scarce. We present the Atmospheric Spatio‐Temporal Reconstruction Network (AtmoST‐Net), which integrates high‐resolution SABER observations with dual‐source reanalysis, combining non‐uniform vertical enhancement and explicit physical forcing factors (as input features) to reconstruct temperature at 20–80 km, improving the mid‐to‐upper atmosphere vertical resolution by 3.6 km. AtmoST‐Net achieves mean absolute error <3K at 20–50 km and improves accuracy by 5.0% in the transition zone (∼65 km). It reliably reproduces the vertical structure of Antarctic sudden stratospheric warming and recovers QBO (Quasi‐Biennial Oscillation) signals in the 50–60 km region that are underrepresented in traditional reanalysis data sets, with the first EOF (empirical orthogonal function) principal component showing correlation with QBO of 0.498 (p < 0.01) and variance contribution reaching 96.06%. Our data set fills critical observational gaps and provides support for atmospheric dynamics research.