🌏 Cross-Domain Foundation Model Adaptation: Pioneering Computer Vision Models for Geophysical Data Analysis

🏢 Computational Interpretation Group (CIG)

Zhixiang Guo¹, Xinming Wu^1*, Luming Liang², Hanlin Sheng¹, Nuo Chen¹, Zhengfa Bi³

School of Earth and Space Sciences, University of Science and Technology of China, Hefei, China

Microsoft Applied Sciences Group, Redmond, WA 98052, United States

Lawrence Berkeley National Laboratory, 1 Cyclotron Rd, CA 94707, USA

📣 News

🛸 The dataset, model, code, and demo are coming soon!

💥 [2024.09.01]: The code has been uploaded.

💥 [2024.08.23]: The paper has been submitted to Arxiv: https://arxiv.org/pdf/2408.12396

💥 [2024.07.23]: Upload the dataset.

💥 [2024.07.07]: Github Repository Initialization.

✨ Introduction

Workflow for adapting pre-trained foundation models to geophysics. First, we prepare geophysical training datasets (1st column), which involves collecting and processing relevant geophysical data to ensure it is suitable for adaption fine-tuning. Next, we load the pre-trained foundation model as the data feature encoder (2nd column) and fine-tune the model to make it adaptable to geophysical data. To map the encoder features to the task-specific targets, we explore suitable decoders (3rd column) for geophysical downstream adaption. Finally, the adapted model is applied to various downstream tasks within the geophysics field (4th column).

🚀 Quick Start

1. Clone the repository

Our code provides demos corresponding to the data mentioned in the paper, including seismic facies, geological bodies, DAS, faults, and craters. You can run them by following the steps below:

First, clone the repository to your local machine:

git clone git@github.com:ProgrammerZXG/Cross-Domain-Foundation-Model-Adaptation.git
cd Cross-Domain-Foundation-Model-Adaptation

2. Install dependencies

pip install -r requirements.txt

3. Download the dataset

Before running the code, you need to download the dataset. You can download the dataset in Zenodo and put them in the data/.

4. Run the code

cd run
bash mla_facies.sh

If you choose to use bash run/mla_facies.sh, please be aware of the dataset path.

🌠 Results

Quantitative Metrics for Downstream Tasks

Mean Intersection over Union (mIoU)

Network	Seismic Facies Classification	Seismic Geobody Identification	Crater Detection	DAS Seismic Event Detection	Deep Fault Detection
Unet	0.5490	0.8636	0.5812	0.7271	0.6858
DINOv2-LINEAR	0.6565	0.8965	0.6857	0.8112	0.6372
DINOv2-PUP	0.6885	0.8935	0.6937	0.8487	0.7088
DINOv2-DPT	0.6709	0.8912	0.6917	0.8672	0.7334
DINOv2-MLA	0.6826	0.8969	0.6949	0.8591	0.7613

Mean Pixel Accuracy (mPA)

Network	Seismic Facies Classification	Seismic Geobody Identification	Crater Detection	DAS Seismic Event Detection	Deep Fault Detection
Unet	0.7693	0.9112	0.6265	0.7865	0.7439
DINOv2-LINEAR	0.8732	0.9374	0.7481	0.9033	0.7519
DINOv2-PUP	0.9102	0.9357	0.7529	0.9210	0.7793
DINOv2-DPT	0.8826	0.9377	0.7462	0.9119	0.7985
DINOv2-MLA	0.8975	0.9383	0.7476	0.9222	0.8195

📦 Dataset

All data is avalable at Zenodo.

Task	Data Sources	Data Size	Training Number	Test Number
Seismic Facies Classification	provided by (SEAM, 2020)	1006 × 782	250	45
Salt Body Identification	provided by (Addison Howard et al., 2018)	224 × 224	3000	1000
Crater Detection	original data provided by CAS, labelled by authors	1022 × 1022	1000	199
DAS Seismic Event Detection	provided by (Biondi et al., 2023)	512 × 512	115	28
Deep Fault Detection	original data provided from field surveys, labelled by authors	896 × 896	1081	269

🔖 Citation

If you find this work useful, please consider citing our paper:

@misc{guo2024crossdomainfoundationmodeladaptation,
      title={Cross-Domain Foundation Model Adaptation: Pioneering Computer Vision Models for Geophysical Data Analysis}, 
      author={Zhixiang Guo and Xinming Wu and Luming Liang and Hanlin Sheng and Nuo Chen and Zhengfa Bi},
      year={2024},
      eprint={2408.12396},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.12396}, 
}

📝 Acknowledgment

This study is strongly supported by the Supercomputing Center of the University of Science and Technology of China, particularly with the provision of Nvidia 80G A100 GPUs, which are crucial for our experiments. We also thank SEAM for providing the seismic facies classification dataset, TGS for the geobody identification dataset, CAS for the crater detection dataset, Biondi for the DAS seismic event detection dataset, and CIG for the deep fault detection dataset.

📮 Contact

If you have any questions about this work, please feel free to contact xinmwu@ustc.edu.cn or zxg3@mail.ustc.edu.cn.