Convolutional Neural Network to Identify IRD in IODP Expedition 382 Sediment Cores X-ray Images

Repository for the Jasper et al. (2024) manuscript in Paleoceanography and Paleoclimatology entitled, "A 3.3-Million-Year Record of Antarctic Iceberg Rafted Debris and Ice Sheet Evolution Quantified by Machine Learning"

This repo has three jupyter notebooks:

1_cnn_build

notebook with the training data set up and the cnn model architecture.

2_label_images

notebook to use the trained cnn model to label either single images or images down splice

3_ird_counter

notebook to go from the IRD-labeled images to a csv file with the depth of each IRD grain identified down the splice

Additional Information

All of the training data and the trained CNN model used in Jasper et al. (2024) has been archived on Zenodo, and is available for download here

Note: The model architecture in 1_cnn_build builds off of the code from Dyer et al. 2021 PNAS and the original model build can be found here

Requirements for running and training the CNN model

CUDA

You need to be running CUDA 10.0 (10.1 might work now) to use tf2.

Environment files

The environment file (tensorflow.yml) contains the package requirements needed to run the notebooks in this repository. Create a local anaconda environment on your machine from the .yml file:

conda env create -f tensorflow.yml

activate that environment:

conda activate tensorflow

and run jupyter:

jupyter notebook

or

jupyter lab