Fast imaging of transient electromagnetic data based on convolutional neural network

The one-dimensional (1D) inversion of Transient Electromagnetic (TEM) data has been well developed, however, it is still difficult to conduct real-time imaging of field survey data. In this paper, we introduce a Convolutional Neural Network (CNN) into TEM data imaging. By training the network to approximate the relationship between the time-domain EM responses and geoelectrical model parameters, the complex inversion process is transformed into a matrix mapping process, realizing the real-time and fast imaging of TEM data. Considering that the traditional imaging algorithms are mostly executed on a point-by-point basis, it is difficult to achieve rapid processing of large survey dataset, we try in this paper to take the coordinate information of the receivers relative to the transmitting source as the network input parameter, which not only makes the algorithm more flexible in the imaging of survey data, but also greatly reduces the number of samplings in the training process. To verify the effectiveness, we first test our algorithm on synthetic data and demonstrate that by adding coordinate information into the network our CNN method has the flexibility of imaging irregular survey points without affecting the imaging accuracy. Finally, we compare our imaging results with the Occam's inversion for a survey dataset to further verify the effectiveness of our imaging method.