Underground three-dimensional displacement monitoring plays an important role in landslide disaster early warning. Compared with surface displacement and underground two-dimensional displacement monitoring, it can provide more accurate early warning information. Previously, the measurement device of underground three-dimensional displacement based on dual mutual inductance voltages was developed, which can collect millimeter-level displacement data. However, how to establish the relationship between mutual inductance voltages and underground three-dimensional displacement still faces huge challenges. In this work, the dual mutual inductance voltages at each inter-axis angle are treated as data maps. Then, they can serve as input to the convolutional neural network-based model (ConvRnet), who possess powerful image data mining capabilities. The experiment results show that the ConvRnet model significantly improves the prediction accuracy compared to the interpolation model and other machine learning models. The maximum errors of horizontal displacement and vertical displacement are 0.42mm and 0.32mm, respectively. The ConvRnet model combined with the data maps of dual mutual inductance voltages provides an effective tool for underground three-dimensional displacement measurement.