Seismic Image Segmentation

This repository presents a segmentation approach for seismic images, based on the fine-tuning of the Segment Anything Model (SAM). The project adapts SAM using the IA3 adapter and leverages combined prompt strategies (bounding box + points) for accurate automatic segmentation of geological structures (e.g., paleovres, reefs). The method achieves improvements in key segmentation metrics (Dice up to 0.60, IoU gain) and maintains real-time feasibility (~1.35 FPS), making it suitable for industrial and research use.

Original paper

🔑 Key Features

🧠 Adapted SAM Model

Utilizes the pre-trained facebook/sam-vit-huge with an IA3 adapter (reducing trainable parameters by ~10%) for efficient seismic segmentation.

🎯 Combined Prompt Strategy

Supports multiple prompt types, including:

Points
Circles
Bounding boxes
Combined (e.g., bounding box + points/circles)
to enhance segmentation precision and generalization.

🛠️ Data Pipeline

Processes seismic datasets through:

2D slice conversion to RGB
Intensity normalization
Binary mask generation

Included datasets:

Real-world data: Salt2D
Synthetic data: sabamrine, paleokart

🧪 Experiment Management

Managed via Hydra for flexible experiment configuration
Integrated with ClearML for tracking, reproducibility, and hyperparameter optimization

🚀 Training & Inference Pipeline

Comprehensive pipelines for:

Training, validation, testing
Automated reporting of IoU and Dice metrics

🧰 Installation

Clone the Repository:

git clone <repository_url>
cd seismic_segmentation

Install Dependencies:

Use a virtual environment before installation.

pip install -r requirements.txt

Download SAM Weights (Optional):

Refer to the segment-anything repository for model checkpoints.

⚙️ Configuration

YAML-Based Setup

Experiment configurations reside in conf/config.yaml and include:

Model Settings:
Model type, adapter (IA3 or LoRA), layer freezing options
Training Hyperparameters:
Epochs, learning rate, batch size, checkpointing frequency
Dataset Settings:
Paths to seismic images and masks, input shape, binary format
Prompt Settings:
Prompt type (bbox+points, etc.), number of points, bbox noise

🔐 ClearML Integration

Set your ClearML credentials via environment variables:

Path: experiments/.env.example

CLEARML_WEB_HOST=https://app.clear.ml/
CLEARML_API_HOST=https://api.clear.ml
CLEARML_FILES_HOST=https://files.clear.ml
CLEARML_API_ACCESS_KEY=your_access_key
CLEARML_API_SECRET_KEY=your_secret_key

Rename the file to .env or set the variables in your shell.

🏃 Running the Training

To launch training:

./run_train.sh

For multi-run experiments:

./run_train.sh --multirun

📁 Data Structure

/data
│
├── Salt2d       # Real-world 2D seismic slices
├── sabamrine    # Synthetic 3D volumes with paleovres
└── paleokart    # Synthetic seismic structures (paleokarts)

🗂 Project Layout

conf/           # Experiment configurations (Hydra)
experiments/    # Training logic and scripts
data/           # Datasets and annotations
run_train.sh    # Entry script for training runs

📌 Notes

Backbone Model: facebook/sam-vit-huge
Adapter: IA3 (efficient fine-tuning)
Tools: Hydra + ClearML for reproducible ML experiments
Prompt Modes: bbox, points, hybrid configurations

🤝 Acknowledgements

This project has benefited from open-source contributions and collaborative experimentation across geoscience and AI communities. We thank everyone who helped shape and test this solution.

petrovdenees
vladimir6567
bc-ru
unlikexd
andreyvostretsov

For technical inquiries or usage questions, refer to the documentation or reach out via the repository’s issue tracker.