AI Classifier for Healthy Coral Reefs: An AI Hackathon Challenge

MARS hosts an internal data science/AI hackathon every year, and for 2023 the challenge was to create a model that could classify healthy coral reefs.

The team consisted of myself and two of my colleagues , Emil and Andrew.

My day-to-day work consists of building computer vision deep learning models and so I was excited to see what I could do in a very different domain:

• The model I submitted had an F1 score of 0.97 which just missed the podium! (4th overall).

• I used an EfficientNetB5 with pre-trained weights and fine-tuned on the coral training set (~2k images) with a ton of tuning and experimentation.

• The basic workflow was:

  1. Interviewing domain experts to learn how humans classify coral reefs.

  2. Exploratory Data Analysis to understand the data (image metadata, distribution of color values, etc.)

  3. Building an initial model with almost no transformations as a benchmark

  4. Hyper-parameter tuning and experimenting with transforms while using MLFlow for experiment tracking.

  5. Model evaluation using FiftyOne. Visually examining mistakes the model was making and looking for common denominators

  6. The above steps were iterated until the best approach was found, and the final model was then trained on ALL of the data before submission.