| Autor: |
Wu, Sihong, Huang, Qinghua, Zhao, Li |
| Předmět: |
|
| Zdroj: |
Geophysical Journal International; Apr2023, Vol. 233 Issue 1, p253-263, 11p |
| Abstrakt: |
Airborne electromagnetic (AEM) method detects the subsurface electrical resistivity structure by inverting the measured electromagnetic field. AEM data inversion is extremely time-consuming when huge volumes of observational data are involved. Forward modelling is an essential part and represents a large proportion of computational cost in the inversion process. In this study, we develop an AEM simulator using deep learning as a computationally efficient alternative to accelerate 1-D forward modelling. Inspired by Google's neural machine translation, our AEM simulator adopts the long short-term memory (LSTM) modules with an encoder–decoder structure, combining the advantages in time-series regression and feature extraction. The well-trained LSTM network describes directly the mapping relationship between resistivity models with transceiver altitudes and time-domain AEM signals. The prediction results of the test set show that 95 per cent of the relative errors at most sampling points fall in the range of ±5 per cent, with average values within the range of ±0.5 per cent, indicating an overall prediction accuracy. We investigate the effects of the distributions of both resistivity and transceiver altitude in the training set on the prediction accuracy. The LSTM-based AEM simulator can effectively handle the resistivity characteristics involved in the training set and yields great sensitivity to the variations of transceiver altitudes. We also examine the adaptability of our AEM simulator for discontinuous resistivity variations. Synthetic tests indicate that the application effect of the AEM simulator relies on the completeness of the training samples and suggest that enriching the sample diversity is necessary to ensure the prediction accuracy, in cases of observation environments dominated by extreme transceiver altitudes or under-represented geological features. Furthermore, we discuss the influence of network configuration on its accuracy and computational efficiency. Our simulator can deliver ∼13 600 1-D forward modelling calculations within 1 s, which significantly improves the simulation efficiency of AEM data. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
