AI classification of jellyfish and plastic from ocean data

Inspiration

Our oceans are changing.

As the largest ecosystem on earth, they face the ever-increasing challenges of plastic pollution and invasive marine jellyfish.

Classifying between them can be challenging and would be essential for the deployment of all-purpose devices that can clean up plastic and exterminate invasive species. Such automation would be invaluable to improve conservation efforts and validate marine surveys, benefitting the world and tackling climate change.

What it does

The program uses a convolution neural network (CNN) to distinguish images of plastic and images of jellyfish from the Jellyfish Object Detection and the DeepPlastic datasets.

How we built it

We began by preprocessing data from a labelled image datasets of marine plastic and jellyfish, preparing them into identical 100x100 images. We passed this into a convolution neural network (CNN), with several layers to combat overfitting. We used a 2/3-1/3 train-test split and accelerated the CNN with the IBM-Z mainframe.

Challenges we ran into

Finding the datasets was the most challenging part, and preprocessing our data to be passed into the CNN required some additional research on the necessary convolutional layers.

Accomplishments that we're proud of

We were concerned that the model may be able to distinguish plastic from jellyfish purely by virtue of them being from different datasets. Thanks to our preprocessing, high validation accuracy, and optimised parameters, this turned out not to be the case.

What we learned

We obtained a deeper understanding of CNNs for classification and how more complicated models could be used for real-life applications.

What's next for AI classification of jellyfish and plastic from ocean data

We would like to try classifying different marine life. This could be done with larger datasets such as FathomNet.

Built With

• ibmz
• python
• tensorflow