A deep learning-based parametric inversion for forecasting water-filled bodies position using electromagnetic method
We design a tunnel electromagnetic joint scan observation system and present a deep learning-based parametric inversion for improved tunnel electromagnetic imaging, designed specifically for tunnel prediction of water filled structures. It utilizes a configuration wherein transmitters scan along the surface while receivers are positioned within the tunnel, employing time-domain and frequency-domain transmitters and a multi-component receiver. The DL model for the first time provides parametric imaging of two different view, forming a self-checking mechanism, which can help constrain the predictions and reduce the non-uniqueness of the inversion. Trained by synthetic data, our system shows impressive adaptability to predict the 3D spatial position of water-filled anomalies and strong robustness in the tunnel environment with metal interference.
Ubuntu 22.04.1 LTS
Python 3.10.6
SimPEG 0.18.1
scipy 1.10.0
numpy 1.23.5
tensorflow 2.10.0
pip install -r requirement.txt
conda install SimPEG --channel conda-forge
For SimPEG, you may refer to more detail at https://docs.simpeg.xyz/latest/content/getting_started/installing.html
Before prediction, you need to download the pre-trained weight model from https://zenodo.org/records/14745131, which contain
UNet_model/FTEM.ckpt.data-00000-of-00001,
UNet_model/FTEM.ckpt.index,
UNet_model/FTEM.h5.
Then place the directory containing the weight model in the same directory as the prediction code.
Then run: python predi.py
The program will read data from directory 3. The data is obtained through scanning observations, allowing for the prediction of the three-dimensional spatial location of the underground water-filled structures.
FTEM_batch_train.ipynb is used to train a new model. The traindata can be download from https://zenodo.org/records/11407019
The Topo directory provides specific examples for terrain-based prediction, gdata.ipynb is used to generate electromagnetic field observation data with terrain, and predi.py is used to predict the data and results will be saved in creted dir by gdata. The prediction results are saved in the "predict" folder.
Gan, L., Tang, R., Li, H., Li, F., & Rao, Y. (2025). A deep learning-based parametric inversion for forecasting water-filled bodies position using electromagnetic method. Computers & Geosciences, 105881.
This project is licensed under the MIT License.