反演(地质)
合成数据
算法
计算机科学
反变换采样
地球物理学
人工智能
地质学
构造盆地
电信
表面波
古生物学
作者
Siyuan Dong,Jian Jiao,Shuai Zhou,Pengyu Lu,Zhaofa Zeng
标识
DOI:10.1109/tgrs.2023.3306980
摘要
Three-dimensional gravity inversion is an effective method for restoring underground density distribution from gravity anomaly data. Conventional regularization inversion has good data fitting, but its inversion model has insufficient model fitting capabilities due to its low-depth resolution. Although data-driven deep learning-based gravity inversion results significantly improve depth resolution and physical property distribution, it is difficult to ensure the data fitting of the inversion results. Accordingly, this study proposes a three-dimensional gravity data inversion based on enhanced dual U-Net framework (EdU-Net) to solve the above problems, making the inversion results have good model and data fitting performance. The proposed EdU-Net consists of two parts: first, training a large generalization pre-trained network Net I, and then quickly generating an enhanced Net II for the target data through fine-tuning. Additionally, this study adds forward-fitting constraints in the framework’s loss function to reduce the problem of large data-fitting errors in traditional data-driven deep learning inversion. The trained Net II inversion result has better model and data fitting accuracy than Net I. Moreover, by comparing the inversion results of synthetic models, this study demonstrates that the EdU-Net method performs better than traditional deep learning. Finally, this method is applied to the measured data of the Gonghe Basin in Qinghai Province, China, and provides a reasonable explanation for the distribution of hot dry rocks.
科研通智能强力驱动
Strongly Powered by AbleSci AI