Magnetization Vector Inversion Method Based on U-Net Neural Network

To address the limitations of magnetization vector inversion, including insufficient inversion accuracy and the non-uniqueness of solutions in magnetic anomaly fields, this paper proposes a U-Net-based inversion framework for magnetization vector estimation. The proposed method constructs a deep learning inversion network that takes magnetic anomaly vector components as inputs, and outputs the three-dimensional magnetization vector. The network architecture adopts a symmetric U-shaped fully convolutional design, incorporating skip connections and normalization strategies to enhance feature extraction and spatial resolution. Dropout layers are introduced to mitigate overfitting, and the mean squared error (MSE) is employed as the loss function to improve inversion accuracy. Experimental results demonstrate that, compared with conventional regularized inversion methods, the proposed approach shows significant advantages in target localization and vector recovery accuracy, verifying its effectiveness and potential for engineering applications in magnetic anomaly inversion.