Abstract: The electromagnetic inversion process of ground penetrating radar(GPR) data can obtain geometric parameters and electromagnetic parameters of underground targets. Aiming at the feature differences between GPR time domain data and frequency domain data in the image domain, this paper designs a deep learning-based GPR dimensional transformation auto-encoder to extract GPR B-scan time domain features and perform uniform processing on GPR data features in time domain and frequency domain. A GPR-EInet electromagnetic inversion framework based on time-frequency fusion data is designed. GPR-EInet is composed of a multi-scale feature extractor and a feature reconstructor. The multi-scale feature extractor realizes the fusion and feature extraction of time-frequency domain data, and improves the anti-noise ability of the proposed network. The feature reconstructor realizes the nonlinear mapping between time-frequency fusion feature and the permittivity distribution of the detection area. 2000 and 200 GPR B-Scan data are produced to train and test GPR-EInet respectively. Simulation results show that GPR-EInet can realize single/dual target electromagnetic inversion and the SSIM between GPR-EInet predicted value and real value is 0.99564 under the conditions of SNR = −10 dB and the deviation between target dielectric constant and background dielectric constant is 50%. These three deep learning network(GPR-EInet, Ünet and PINet) are used to conduct electromagnetic inversion of the simulated data respectively. The results show that the anti-noise performance of GPR-EInet is better than that of PINet and Ünet. Electromagnetic inversion experiments are also carried out on on-site GPR data. Compared with using time domain data and frequency domain data alone, the time-frequency fusion strategy of GPR-EInet improves the electromagnetic inversion quality and anti-noise ability.