Classifying Images of Tide-Zone Invertebrate Species Through Neural Network Technology

Noah Jaffe, a recent graduate of the Metis data science bootcamp, has completed a fascinating project that leverages deep learning to classify large groups of intertidal invertebrates. The project, which was inspired by Jaffe's Master's in marine biology where he studied sea star wasting disease, marks a significant step in the application of neural networks to the task of identifying small, intertidal invertebrates.

The Project Setup

To begin, Jaffe constructed the necessary configuration, object, and training text files for his model. He chose to use the YOLO v3 convolutional neural network, a popular choice for object detection tasks. To train his model, he followed a tutorial to utilise the GPU functionality of Google Colab.

Training the Model

Jaffe trained his model using Google Colab, with a GPU, for approximately three days. However, he encountered challenges with Google Colab interrupting his processes for long periods of time. To mitigate this issue, he wrote code to save his model output every 100 iterations.

The training data for the project consisted of images of anemones, barnacles, bivalves, crabs, nudibranchs, and sea stars. Jaffe used LabelImg to label his training images, and a Google Chrome plugin called Download All Images to gather his training images. He used the Darknet NN framework to train his model, a framework written in C that allows for working with models in a notebook.

Testing and Optimisation

Jaffe tested his model using OpenCV, following a tutorial, and defined testing functions to test his model. His classifier successfully identified intertidal invertebrates, but was not perfect, with crabs being especially difficult due to their various positions. Jaffe views his project as a proof of concept for the application of neural networks to the task of classifying small, intertidal invertebrates, and believes that further training and hyperparameter tinkering could improve the model.

Final Deliverable

The final deliverable for the project was a Streamlit app, where users can upload images of intertidal invertebrates, and the app will return the predicted classification. Jaffe included a picture of a classifier example in the "Project_5_final_pres.pdf" file in the GitHub repo.

Jaffe's project represents a successful first step in constructing a tool to identify intertidal invertebrates, and opens up exciting possibilities for future research in this area.