PATSA-BIL: Pipeline for Automated Texture and Structure Analysis of Borehole Image Logs

This repository contains the code and data for PATSA-BIL, a robust pipeline for automated analysis of borehole image logs. The project provides preprocessing, classification, and segmentation pipelines for resistivity image log data.

---

Highlights

• We propose PATSA-BIL, a robust pipeline for analyzing resistivity image log data.
• PATSA-BIL achieved up to 90% accuracy in texture classification using ViT models.
• PATSA-BIL with SLIC+DBSCAN outperformed baseline DINOv2 by up to 16% (mAP).
• A Wilcoxon–Holm analysis confirmed the significance and stability of our approach.

Citing

If you use PATSA-BIL in your research, please cite:

@article{Souza2025,
  title = {{PATSA}-{BIL}: Pipeline for automated texture and structure analysis of borehole image logs},
  volume = {278},
  ISSN = {0957-4174},
  url = {http://dx.doi.org/10.1016/j.eswa.2025.127345},
  DOI = {10.1016/j.eswa.2025.127345},
  journal = {Expert Systems with Applications},
  publisher = {Elsevier BV},
  author = {Souza, André M. and Cruz, Matheus A. and Braga, Paola M.C. and Piva, Rodrigo B. and Dias, Rodrigo A.C. and Siqueira, Paulo R. and Trevizan, Willian A. and de Jesus, Candida M. and Bazzarella, Camilla and Monteiro, Rodrigo S. and Bernardini, Flavia C. and Fernandes, Leandro A.F. and de Sousa, Elaine P.M. and de Oliveira, Daniel and Bedo, Marcos},
  year = {2025},
  month = jun,
  pages = {127345}
}

---

Prerequisites

Install dependencies using pip or conda:

• pip:

  pip install -r requirements.txt

• conda:

  conda env create -f environment.yml
  conda activate <environment_name>

Configure environment variables:

python setup.py

---

Library usage

Data structure

In the root folder (imagelog) there is the data folder, which contains three mandatory folders: raw, interim, and processed.

• data/raw: Original, unprocessed data.
• data/interim: Minimally processed data (e.g., normalization, invalid portion removal).
• data/processed: Fully preprocessed data for each project.

The raw folder contains the original data, without any changes. The interim folder contains the data after minimal processing (removal of invalid portions, normalization, saved as images). The processed folder contains subfolders, with each containing the data after the specific preprocessing pipeline of a project.

Project Structure

• projects/: Contains project folders.
  • Each project has experiments/ and preprocesses/ subfolders.
  • Configuration files: <experiment_name>.json and <preprocess_name>.json.

In the root folder, there is the projects folder, containing folders that represent different project specifications. Each project contains experiments and preprocesses subfolders, with each subfolder having a JSON file (<experiment_name>.json/<preprocess_name>.json) containing the configuration of the experiment/preprocessing steps and a list of datasets to be processed.

---

Data Operations

1. Reading ImageLog in CSV

To convert a CSV to images, use the imagelog_csv_loading.py script, with the input folder located in raw and the name of the folder to be generated in interim:

python scripts/imagelog_csv_loading.py csv_to_images --input_data_folder "<DATASET_RAW>" --output_data_folder "<DATASET_INTERIM>"

2. Data preprocessing

Data preprocessing for a project is done through the following command:

python experiment_scripts/classification.py preprocess --project_name "<project_name>" --preprocess_name "<preprocess_name>" --override_preprocess

3. Synthetic Data Generation

To generate synthetic imagelog data, use the generate_data command from the imagelog_dataset_generator.py script.

python scripts/imagelog_dataset_generator.py generate_data --dataset_name "SYNTH_TEST" --batch_size 5 --number_of_tiles 5 --apply_colormap

• dataset_name: Output folder in interim
• batch_size: Number of images per batch
• number_of_tiles: Total images per pattern (multiple of batch_size)
• apply_colormap: Use YlOrBr color scale if set

---

Classification Model Operations

Training & Testing

python experiment_scripts/classification.py fit --project_name "tutorial_project" --experiment_name "<experiment_name>" --preprocess_name "<preprocess_name>"
python experiment_scripts/classification.py test --project_name "tutorial_project" --experiment_name "<experiment_name>" --preprocess_name "<preprocess_name>"

Prediction

python scripts/classification.py predict --project_name "tutorial_project" --experiment_name "<experiment_name>" --preprocess_name "<preprocess_name>"

KFold Optimization

To perform the KFold optimization experiment to evaluate the performance of a model, use the kfold and full_kfold commands from the classification.py script along with the flags specifying the project and the experiment file name.

python scripts/classification.py kfold --project_name "tutorial_project" --experiment_name "<experiment_name>" --preprocess_name "<preprocess_name>"

---

Segmentation Model Pipelines

Differently from the classification pipelines, we implemented segmentation pipelines in a single script for each method (along with Jupyter notebooks for each experimental pipeline). The segmentation pipelines are located in the experiment_scripts folder: mask_rcnn.py, sam.py, slicdbscan.py. Each segmentation pipeline expects data annotated with masks (we used the LabelStudio tool for labeling), which are used to train the model. The data is expected to be in the processed/<project_name>/<preprocess_name>/<datasource_name>/images folder, with JSON annotation files for each image in the processed/<project_name>/<preprocess_name>/<datasource_name>/labels subfolder.

Segmentation scripts are in experiment_scripts/:

• mask_rcnn.py
• sam.py
• slicdbscan.py

Data requirements:

• Images: processed/<project_name>/<preprocess_name>/<datasource_name>/images
• Masks (JSON): processed/<project_name>/<preprocess_name>/<datasource_name>/labels
• Annotation tool: LabelStudio

---

License

Distributed under the GNU GENERAL PUBLIC LICENSE License. See LICENSE for details.

---

Contact

For questions or contributions, please open an issue or contact the authors.