Vision Transformer-Based H-k Method(HkViT) for Predicting Crustal Thickness and Vp/Vs Ratio from Receiver Functions ——Final project for computer vision (04835030) at PKU
Our report are posted in the Respository, our video will soon be released here
Crustal thickness (H) and crustal Vp/Vs ratio k are fundamental parameters for regional geology and tectonics.The teleseismic receiver function (RF) is the response of the Earth structure below a seismometer to an incident seismic wave. It is commonly used to determine major interfaces of the Earth, including the above two parameters.
Up to now, a deep learning-based H-k Method (HkNet) has been proposed with higher accuracy and more stable results comparing to H-k-c method. However, there are still quite large room for HkNet to improve its accuracy and robustness. Here we propose a new method which uses Vision Transformer (ViT) to estimate H and k.
Our model can be divided into two parts. The first part is set to denoise the receiver functions, while the second part uses ViT to predict H and with denoised RFs being its input. Synthetic data tests and real data both show that our new method obtains great accuracy and robustness.
Feel free to contact us through email if you have any questions!
- Python 3.8
- PyTorch
- torchvision
- timm
- vit_pytorch
- tqdm
- pandas
Install dependencies with the following command:
pip3 install -r requirements.txt
you might need to adapt root for data in Dataset.py
To train our model and Denoise net:
python main.py
you could change config through parameters of train_reg、train_denoise in main.py