.png){kind=icon}
.png){kind=icon}
Return
CNN-based transfer learning for self-potential inversion
.png){kind=icon}
PRE
AI
DOI：10.1007/s12145-025-01951-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Self-potential (SP) inversion is an effective geophysical technique for metal mineral exploration. Traditional deep learning-based inversion methods are prone to overfitting and may fail to produce reliable results in scenarios with limited samples. To address this challenge, a two-dimensional convolutional neural network (2DCNN)-based transfer learning (TL) framework TL-2DCNN is proposed to interpret SP anomalies as regular polarized bodies. We evaluate its capability using synthetic data and conduct comparative analyses with other deep learning methods. To further test the efficacy of the TL-2DCNN in complex conditions, limited-sample experimental data are used for transfer network training. The inversion results of two test datasets illustrate that the TL-2DCNN can deliver reliable parameter estimations even when pre-training datasets are derived from regular polarized bodies, confirming the feasibility of the transfer training strategy. Overall, the superior performance of the TL-2DCNN framework establishes it an effective tool for SP inversion under sample-limited conditions in practical exploration.
Keywords:
Self-potential
Transfer learning
2D CNN
Inversion
Limited samples
Journal
IF:
3
Papers:
634
Citations:
3.3K