Explainable Detection of Microplastics Using Transformer Neural Networks

This repository provides supplementary material for the following publications (would appreciate a citation of our work if you find the tool useful)

M. Barker, M. Willans, D-S. Pham, A. Krishna, M. Hackett. Explainable Detection of Microplastics Using Transformer Neural Networks, in Proceedings of the Australasian Joint Conference on Artificial Intelligence (AJCAI), Perth December 2022.

Available in this repository:

• Reflectance micro-FTIR spectral data for standard and marine polymers
• Python code that implements the model

Setup and Execution

This code runs using pytorch-gpu which you can download from pytorch.org/.

Assuming pytorch is installed you can run the program with:

python3 run.py

Configuring

All of the variables that control the models hyperparameters are in run.py. I last tested this code on a GTX 1080ti, so you may be able to change certain hyperparameters.

Currently the dataloaders load the micro-FTIR datasets that contain polyethylene and polypropylene in the no_fp directory. There is however an alternative option to train the model with filter paper samples. This can be done by using the csv files in the with_fp directory. Otherwise, you can extract the entire datasets in the marine_polymers.csv and standard_polymers.csv in the data directory.

You will need to change the global variable LABEL_DICT if you wish to use the filter paper samples. This can be done by changing:

LABEL_DICT = {'PP': 0,'PE':1}

to

LABEL_DICT = {'PP': 0,'PE':1,'FP':2}

Figures

Reflectance Micro-FTIR Workflow

Raw Marine Polymer Data and First Order Difference

Best Performing Validation and Test Results

Baseline Model Comparison