Seismic Segmentation Methods Benchmark

This project was heavily inspired by the works of Alaudah et al. in 'A Machine Learning Benchmark for Facies Classification' [Paper][Code].

Our goal is to propose a benchmark method for comparing the results of different deep learning approaches for the task of facies segmentation in seismic volumes. We provide the implementations of three common encoder-decoder models (SegNet, U-Net and DeconvNet) as well as three open-source labeled datasets (F3 Netherlands, New Zeland Parihaka and Nova Scotia Penobscot) to serve as baselines. This code is structured in a way as to easily allow for new architectures and datasets to be incorporated into the benchmark, as a means to encourage future research in the field.

Instaling the Required Packages

This project was coded and tested using Python 3.12.2, but older versions of Python 3 should work without issue. To install all the depedencies required, execute the following command within the main project folder:

pip install -r requirements.txt

You might want to use older versions of these packages to match your Python version as well.

Downloading the Data

All three of the datasets listed above, as well as their respective annotations, were preprocessed and compressed into datasets.tar.xz as .npy files. They can be obtained through the gdown package by running the following commands:

pip install gdown
gdown 1KtUr3dbcf_BBWKDkRWTil8f89CaW8e9W

To use the data, move the .tar.xz file to the desired folder and extract it by running the following command:

tar -xJf datasets.tar.xz

Command Line Arguments

To train a model, simply run a command like the example:

python3 segment_seismic.py -t -a unet -d path/to/data.npy -l path/to/labels.npy

To test a previously stored model, simply drop the -t flag and specificy the path to a .pt file with -m:

python3 segment_seismic.py -a unet -d path/to/data.npy -l path/to/labels.npy -m path/to/model.pt

While architecture, data_path and labels_path are the only required arguments, there are also several optional arguments to specify hyperparameters. The complete list of arguments, with their descriptions and default values, is given below:

Argument	Description	Default
-a, --architecture	Choose a model to train with. Options are segnet, unet and deconvnet.
-d, --data-path	Path to the seismic volume file in .npy or .segy format.
-l, --labels-path	Path to the labels file in .npy format.
-t, --train	Whether to train a model from scratch. If left on False, the path to a previously stored model must be provided.	False
-b, --batch-size	Size of the training batch.	16
-D, --device	Choose on which GPU to train on. Defaults to the CPU if the device isn't available.	cuda:0
-v, --cross-validation	Whether to train the model using 5-Fold Cross Validation.	False
-f, --test-fold	Choose which of the five chuncks of contiguous slices will be used for testing the model. If --cross-validation is True, this will be ignored and all chuncks will be used as folds.	1
-L, --loss-function	Loss function to use. Currently limited to cel (Cross Entropy Loss).	cel
-o, --optimizer	Optimizer to use. Options are adam and sgd (Stochastic Gradient Descent).	adam
-r, --learning-rate	Learning rate to use during training.	1e-4
-w, --weight-decay	Whether to use weight decay (L2 regularization) in the optimizer to prevent overfitting. A value of 0 indicates no decay.	1e-5
-W, --weighted-loss	Whether to use class weights in the loss function. The weights are used to assign a higher penalty to misclassifications of minority classes.	False
-e, --n-epochs	Number of epochs during training. The actual number might be lower since early stopping is on by default.	30
-O, --orientation	Choose an orientation for slices of the seismic cube to be sampled. Options are in for inlines and cross for crosslines.	in
-F, --faulty-slices-list	Path to a .json file with indexes of faulty slices in the data that contain artifacts. These slices will be discarded from the volume before training.	None
-s, --swap-train-test	Swaps the train and test sets to train the model using less data, that is, only one fold will be used for training while the remaining four will be used for testing.	False
-m, --model-path	Path to a .pt file containing model weights from a previous training. This path must be provided if --train is set to False, as no model will be generated for testing. If a path is provided AND this flag is true, then training will resume from the stored model.	None
-p, --results-path	Directory for storing results. This will generate a .json file containing the metrics and a folder containing the predictions from the test set in .png format. If --train is True, a .pt file will also be created to store model weights.	results