Inversion Method of a Highly Generalized Neural Network Based on Rademacher Complexity for Rough Media GATEM Data
Inversion Method of a Highly Generalized Neural Network Based on Rademacher Complexity for Rough Media GATEM Data
复制标题
基于Rademacher复杂度的高度广义神经网络粗糙介质GATEM数据反演方法
DOI：
10.1109/tgrs.2022.3153686
复制
发表时间：
2022-01-01
影响因子：
8.2
通讯作者：
Wu, Qiong
中科院分区：
文献类型：
--
作者：
Ji, Yanju;Zhang, Yuehan;Wu, Qiong
地源机载时域电磁（GATEM）方法是一种有效的电磁勘探技术。实际地质介质具有粗糙性；然而，目前的反演方法大多基于均匀介质，只提取电阻率信息。本文采用神经网络方法提取粗糙介质的电阻率和粗糙度两个参数。神经网络的结构参数选择没有固定的公式，往往与经验有关。如果结构参数选择不当，神经网络难以收敛到目标精度。为了实现GATEM数据的高精度反演，本文引入了Rademacher复杂度来限制泛化误差，提高神经网络的泛化能力。首先，建立了粗糙介质的GATEM响应、电阻率和粗糙度的样本集。其次，构造了一个完全连通的神经网络结构，并利用Rademacher复杂度得到了一个高度一般化的神经网络。然后通过训练建立映射关系，利用神经网络方法反演电阻率和粗糙度。利用初始神经网络和高度广义神经网络对典型地质模型粗糙介质的GATEM响应进行了反演。基于Rademacher复杂度的高度广义神经网络的结果更接近真实模型。将该方法应用于安徽朱xianzhuang地区GATEM野外资料，结果与地质资料吻合较好。
The ground-source airborne time-domain electromagnetic (GATEM) method is an effective electromagnetic exploration technology. The actual geological medium has rough characteristics; however, the current inversion methods for GATEM data are mostly based on homogeneous medium and extract only resistivity information. In this article, a neural network (NN) is served as extracting the two parameters of resistivity and roughness for rough medium. The structural parameter selection of NN has no fixed formula and is often related to experience. The NN has difficulty converging to the target accuracy if the structural parameters are not selected properly. To realize high-precision inversion of GATEM data, this article introduces Rademacher complexity to limit the generalization error and improve the generalization ability of the NN. Above all, a sample set of the GATEM response, resistivity, and roughness of the rough medium is established. In the next place, a fully connected NN structure is constructed, and a highly generalized NN is obtained by using Rademacher complexity. Then the mapping relationships are established through training, and the NN method is served as inverting the resistivity and roughness. The initial NN and the highly generalized NN are used to invert the GATEM response of rough medium for typical geological models. The results of the highly generalized NN based on Rademacher complexity are closer to the real models. The method is applied to the GATEM field data in Zhuxianzhuang, Anhui Province, China, and the results are consistent with the geological data.