<p>Controlled-source electromagnetic method (CSEM) has been widely applied in engineering and environmental surveys, resource and energy exploration, as well as geological disaster detection. However, due to the increasingly human noises, CSEM data are inevitably subjected to electromagnetic noise, which severely affect the detection results. To address this issue, we propose an improved vision transformer (IVIT) deep learning denoising network to suppress cultural noise, and use wide-field electromagnetic (WFEM, a kind of CSEM) data as an example to validate the effectiveness and superiority. First, typical high-quality CSEM data are selected and a series of simulated noises are added to create a sample library. Second, the well-prepared sample library is used to train the IVIT network. Finally, the well-trained model is employed to perform a one-step denoising operation on the noisy CSEM data to obtain high-quality data. Comparative experiments are conducted with denoising convolutional neural network (DnCNN), residual networks (ResNet), residual DnCNN (ResDnCNN), ResDnCNN combined with shift-invariant sparse coding (ResDnCNN-SISC), U-Net networks, and long short-term memory networks (LSTM). The proposed method effectively removes white noise, pulse noise, and square wave noise, and improves the signal-to-noise ratio (SNR) by approximately 20 dB. Compared with the competitive methods, it has obvious advantages. Analysis of CSEM data from Sichuan Province, China, shows that the data processed by the proposed method results in smoother apparent resistivity curves. In summary, the proposed denoising method can be used to suppress the strong noise of CSEM data, which is helpful for subsequent research of inversion.</p>