With the growing use of airborne platforms in Earth observation, accurate tropospheric delay corrections across various altitudes have become essential. Most existing tropospheric delay models are referenced to the Earth’s surface and rely on analytical closed-form vertical adjustments to approximate delays at user heights. However, these analytical models often fail to capture the complex vertical variations in atmospheric conditions.

To address this limitation, we developed a novel approach leveraging deep neural networks (DNN) to reconstruct global three-dimensional zenith hydrostatic delay (ZHD) and zenith wet delay (ZWD) derived from numerical weather models (NWM). Using the ERA5 pressure-level product (Hersbach et al., 2020) for model training, the DNN refines predictions by correcting the residuals of an analytical third-order exponential model (EXP3). This hybrid method takes advantage of the non-linear fitting capabilities of DNN, significantly enhancing the accuracy of vertical tropospheric delay corrections up to 14 km above the Earth’s surface. The model achieves an average precision of 0.4 mm for ZHD and 0.8 mm for ZWD, reducing root-mean-square (RMS) errors by 63% and 36%, respectively, compared to EXP3.

This dataset includes the EXP3 model, structured on a 1° × 1° global grid at four synoptic times daily (00:00, 06:00, 12:00, and 18:00 UTC) for the period 2019–2022. Additionally, it provides the corresponding DNN model to correct errors in the EXP3 predictions. It is important to note that the model is designed for altitudes ranging from the Earth’s surface up to 14 km.