Deep learning reduction to the pole method constrained by initial model for low latitude magnetic anomalies

Reduction-to-the-pole (RTP) is a vital basic work of magnetic survey data processing. The operation of RTP at low latitude is unstable. Deep learning is an effective data-driven method to solve unstable problems. However, the existing potential field data processing based on deep learning only takes the field response as input. It lacks the universality of off-scene sample data, which often leads to the insufficient generalization performance of prediction results of the deep learning network. Therefore, in order to overcome this problem and further improve the accuracy of RTP at the low latitude, inspired by the "knowledge & data" co-driven deep learning network structure, we propose a fully convolutional RTP network structure for low latitude magnetic anomalies based on initial model constraints. The input of the network structure consists of two parts, one is the low latitude magnetic anomaly, the other is the initial model, and the output is the magnetic anomaly in the direction of vertical magnetization. The magnetic anomaly data of training and testing were generated by fast forward modeling in the spatial domain based on grid point lattice function, and the initial model was obtained by stable frequency conversion for low latitude magnetic anomaly. In addition, 5% Gaussian noise is added to 10% random samples of the data set to strengthen the robustness of fully convolutional network structure. Theoretical model tests verify the effectiveness, accuracy and robustness of the proposed method. Finally, the proposed method is applied to magnetic data of original data, and good results are obtained.