Towards User-Guided Seismic Facies Interpretation with a Pre-trained Large Vision Model

Architecture

Dataset

• All data used for training, validating and testing can be downloaded here.

• The data directory should have this structure:

  data   
    ├── sa.yaml
    ├── images
    │    ├── train
    │    │    ├── image_inline_300.jpg
    │    │    └── ...
    │    ├── val
    │    │    ├── test1_image_inline_100.jpg
    │    │    └── ...
    │    └── test
    │         ├── test2_image_inline_100.jpg
    │         └── ...
    └── labels
    |    ├── train
    |    │    ├── image_inline_300.txt
    |    │    └── ...
    |    ├── val
    |    │    ├── test1_image_inline_100.txt
    |    │    └── ...
    |    └── test
    |    |    ├── test2_image_inline_100.txt
    |    |    └── ...
    |    └── test.cache
    |    |
    |    └── train.cache
    |    |
    |    └── val.cache   
    └── masks
        ├── train
        │    ├── image_inline_300.png
        │    └── ...
        ├── val
        │    ├── test1_image_inline_100.png
        │    └── ...
        └── test
            ├── test2_image_inline_100.png
            └── ...

Requirements

Install requirements

pip install -r requirements.txt
pip install git+https://github.com/openai/CLIP.git

• +NVIDIA GPU + CUDA CuDNN
• +Linux (Ubuntu)
• +Python v3.10

Model Checkpoints

FaciesSAM has four variants:

• FaciesSAM-s-320.pt: trained on image size of 320 and has 12M parameters
• FaciesSAM-s.pt: trained on image size of 640 and has 12M parameters
• FaciesSAM-x-320.pt: trained on image size of 320 and has 72M parameters
• FaciesSAM-x.pt: trained on image size of 640 and has 72M parameters

All model checkpoints can be accessed here. Note: FaciesSAMv2.pt is based on YOLOv9c-seg

Get the validation/test instance segmentation results

Open a terminal and run

python test_instance.py\
        --model_path 'models/FaciesSAM-x.pt'\ # model trained on image size of 640
        --cfg 'data/sa.yaml'\
        --split 'val'\
        --img_sz 640 # Image size can either be 320 or 640

This automatically create a new directory called run. Navigate to see results

Get the validation/test semantic segmentation results

Open terminal and run

python test_semantic.py\
        --model_path 'models/FaciesSAM-x.pt'\
        --data_path 'data/sa.yaml'\
        --split 'test'\
        --img_sz 640  # Image size can either be 320 or 640

This automatically create a new directory called run. Navigate to see results

How to train FaciesSAM on your dataset

Open terminal and run

python train.py \
        --model_path 'models/FaciesSAM-x.pt' \
        --cfg 'data/sa.yaml' \
        --aug True \
        --num_freeze 0 \
        --epochs 100 \
        --bs 16 \
        --img_sz 640 # Image size can either be 320 or 640

This automatically create a new directory called run. Navigate to see results

Interact with seismic image through prompt action

Navigate to the app directory to see how FaciesSAM+CLIP can allow you interact with seismic images

Citation

@article{atolagbe2025faciessam,
  title={Towards User-Guided Seismic Facies Interpretation with a Pre-trained Large Vision Model},
  author={Joshua Atolagbe and Ardiansyah Koeshidayatullah},
  journal=IEEE Access,
  year={2025}
}

Credit

The codes in this repository are based on FastSAM and Ultralytics