AW3D30 DEM data is one of the most widely used basic geographic information data， and its accuracy directly affects the reliability and rigor of a series of derivative products. Therefore， the accuracy validation and improvement of AW3D30 DEM data has always been a research hotspot.. However， conventional high-precision verification data are difficult to obtain and expensive to apply in a wide range of research areas. With global coverage and sub-meter elevation accuracy， ICESat-2 data can provide reliable reference data source for AW3D30 DEM data accuracy validation and improvement. Therefore， this paper takes Henan Province as the study area， and uses ICESat-2 data to validate the elevation accuracy of AW3D30 DEM from the perspective of slope， aspect， geomorphic type and land use type and proposes the Random Forest-Long Short Term Memory Network（RF-LSTM） hybrid model to improve AW3D30 DEM. The results show that the elevation accuracy of AW3D30 DEM decreases with the increase of slope， elevation and topographic relief. The slope direction has less influence on AW3D30 DEM’s elevation accuracy， and the error distribution has no obvious regularity. The accuracy is higher in bare land and cultivated land， and worse in woodland land. The RF-LSTM hybrid model can significantly reduce the mean absolute error and root mean square error of AW3D30 DEM， improve the accuracy of AW3D30 DEM， and provide a reference for the establishment of other DEM data improvement models.