<p>The Global Navigation Satellite Systems (GNSS) tropospheric tomographic technique has been popularly applied to the reconstruction of three-dimensional (3D) water vapour density (WVD) field. Such 3D WVD fields are regarded as a powerful basis of data source in the forecasting of extreme weather events. However, the inherent ill-posed problem in a tropospheric tomographic system needs to be solved properly, and the utilization of virtual signals in a tomographic system is a promising solution. In this study, an improved GNSS tropospheric tomographic method using both real GNSS signals observed and additional virtual signals from the virtual-signals model (VSM) was proposed. The VSM for a target tomographic epoch was constructed using the back-propagation neural network technique and real GNSS signals of the same epoch. One of the clear benefits is that the VSM could generate virtual signals for any elevation and any azimuth angles within the tomographic region for the epoch, i.e. the virtual signals could cross through all voxels in the tomographic region, if sufficient virtual signals were generated. In addition, the parameters of the VSM for different tomographic epochs were also adaptively selected. The VSM in the tomographic experiment was tested using data from the Hong Kong region during the period of DOY 214–244, 2020. Results showed that the root mean square errors (RMSE) of WVD from both improved (GNSS and virtual signals) and traditional (GNSS signals only) tomographic models were 1.14/1.90&#xa0;g/m³ when radiosonde data were used as the reference, and 1.00/1.96&#xa0;g/m³ when ERA5 data were used as references. These results suggest that the proposed VSM can effectively improve tropospheric tomographic results.</p>