A Deep Learning Estimation of the Earth Resistivity Model for ATEM

The Transient Electromagnetic (TEM) method is a geophysical method based on the law of electromagnetic induction. The traditional TEM method has low efficiency because of the ground observation. The airborne TEM (ATEM) was introduced to overcome this limitation. However, due to the flight-movement observation, the useful signal in ATEM data is relatively weak, while the characteristics of noises are complex. So far the denoising process generally consists of many steps, and each step deals with one kind of noise, so the requirements placed on the processing personnel will be very high. After the denoising is completed, the data will be brought into the inversion to obtain an estimate of the earth resistivity model. The inversion generally cannot further distinguish whether there are residual noises in the processed data. Therefore, the reliability of the estimated earth resistivity model will be seriously affected due to the probable disadvantages in the denoising process. Here, we uploaded the data and code used to train a SAE network for estimating earth resistivity from noisy ATEM data. For training of the SAE network, a common solution is to train each sub-network separately, and then form the entire network, in particular: (1) the sub-network 1 will be trained firstly as a separate denoising autoencoder, and at this time, the sub-network 2 and the sub-network 3 will be not considered. (2) After the convergence of sub-network 1 is achieved, the joint structure of sub-network 1 and sub-network 2 will be trained. At this time, the parameters of sub-network 1 will be locked, and only sub-network 2 will be updated. (3) After the convergence of sub-network 2 is achieved, sub-network 3 will be trained in a similar way. When the three sub-networks converge respectively, the training of the SAE can be completed after several joint training of the entire structure, so that the three sub-networks are finally integrated.