.png){kind=icon}
.png){kind=icon}
Return
RTM Gravity Forward Modeling Using Improved Fully Connected Deep Neural Networks
.png){kind=icon}
OA
AI
DOI：10.1109/TGRS.2024.3456812.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The high-frequency gravity forward modeling relying on the residual terrain modeling (RTM) technique is essential for gravity data processing, fine gravity field modeling, geophysical inversion, and so on. However, classical gravity forward modeling methods face challenges such as series divergence and inefficient computation. To improve the computation efficiency, a novel approach using fully connected deep neural network (FC-DNN) for RTM terrain gravity field modeling is introduced in this study. By employing mean squared error (MSE) as the loss function, the method directly learns the mapping between terrain and gravity anomaly to predict RTM terrain gravity anomaly at any elevation, significantly enhancing computational efficiency. In addition, to boost the network's generalization capability, a novel terrain information fusion regularization method is utilized to create an Improved FC-DNN with a refined loss function. The accuracy, computational efficiency, and generalization performance of FC-DNN and Improved FC-DNN are evaluated and compared in the Wudalianchi volcanic region and the Himalayas. The findings reveal that determined RTM terrain gravity fields based on both FC-DNN and Improved FC-DNN meet the mGal-level accuracy in these regions, with a remarkable 10 $000\times $ increase in computational efficiency compared to the classical Newtonian integration method. The Improved FC-DNN exhibits superior generalization ability, with accuracy enhancements ranging from 7% to 21% compared with FC-DNN.
Keywords:
Gravity
Computational modeling
Artificial neural networks
Data models
Computational efficiency
Analytical models
Neurons
Fully connected deep neural network (FC-DNN)
Improved FC-DNN
loss function
residual terrain modeling (RTM) gravity field
transfer learning
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W