Integrating neighboring structure knowledge into a CNN-Transformer hybrid model for global open-access DEM Correction using ICESat-2 altimetry

<b>Accurate global digital elevation models (GDEMs) are essential for various geoscience applications. However, the accuracy of GDEMs in vegetated mountainous regions is relatively low due to substantial topographic relief and the penetration limitations of data acquisition techniques. The Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) acquires high-precision and high-density elevation measurements along its ground tracks on a global scale, offering reliable reference data for GDEM corrections. However, previous GDEM corrections using ICESat-2 altimetry data primarily focused on pixel-by-pixel vertical elevation corrections, often neglecting inter-pixel neighboring structure information, which is crucial for terrain modeling and analysis, particularly in areas with high relief. Therefore, this study incorporates not only ICESat-2 precise elevation observations and but also its along-track neighboring structure knowledge into a convolutional neural network (CNN)-transformer hybrid model, termed NSCT, to correct GDEMs in global scale. The 1 arc-second Copernicus DEM was selected as the target DEM for correction due to its demonstrated superior accuracy. The proposed NSCT model, trained on a diverse range of globally distributed areas with varying topography and vegetation, was evaluated using ICESat-2, global control points, and high-resolution DEMs. Its performance was compared against eight currently most used DEM correction models and publicly available corrected GDEM products, including FABDEM, FathomDEM, and GEDTM30. Correction results demonstrated that the NSCT model generally improved Copernicus DEM accuracy by 66.34%, and outperformed existing correction models in both vertical elevation and neighboring structure assessment across diverse topographic and vegetation conditions. Furthermore, validation using ICESat-2 data and high-resolution DEMs outside the training area, as well with its application to SRTM and AW3D30 GDEM, demonstrated that the NSCT model exhibited superior transferability capability and consistently outstanding performance. This study is the first to integrate the precise elevation observations of ICESat-2 with along-track neighboring structure knowledge into GDEM corrections, offering valuable insights for future research in terrain modeling and analysis.</b>