AI/ML Showcase Portfolio

AI/ML models, tools, and deployment examples focused on marine ecosystem applications .

Full AI/ML Portfolio at this Link: https://github.com/MichaelAkridge-NOAA/open-science-ai-showcase

 
 

Coral Health (Bleaching) Classifier (ViT)

Using a Patch Based Vision Transformer Model Architecture (google/vit-base-patch16-224)

Description:

  • Developed and trained a coral bleaching image classification model using a Vision Transformer (ViT) architecture to distinguish between healthy coral and bleached coral.

Model Architecture:

  • Leveraged and fine-tuned a pre-trained Google ViT-base-patch16-224 model.

  • Fine-tuned only the classifier head while freezing the backbone to preserve learned features and reduce training time.

Data Processing:

  • Processed point-based annotations from NOAA/CoralNet to create AI-Ready training dataset.

  • Generated cropped/patch-based images aligned with annotated points to ensure consistent input sizes and accurate label matching.

  • Organized data into structured folders compatible with classification pipelines (train/validation/test splits).

  • Employed the Hugging Face Datasets library to load and preprocess images.

  • Converted images to RGB format and applied tokenization with AutoImageProcessor for ViT compatibility.

  • Implemented efficient data transformation and collation functions to manage pixel values and label mappings.

Training & Evaluation:

  • Defined training parameters using the TrainingArguments class, optimizing for 100 epochs with batch sizes of 16 and a learning rate of 3e-4.

  • Conducted evaluation at each epoch to monitor performance using an accuracy metric from the evaluate library.

  • Achieved consistent model convergence while limiting overfitting via early stopping and best-model loading strategies.

Results & Deployment:

  • Achieved high classification accuracy in differentiating coral health states.

  • Exported and saved both the trained model and processor for future deployment or inference.

  • Future Model training readily deployable using developed pipelines.

  • Deployed to Demo page using python streamlit and gradio

 

Real-Time(On Edge) Coral Heath (Bleaching) Classifier

Trained to classify coral bleaching conditions using the YOLO11n architecture on imagery from NOAA-PIFSC Ecosystem Sciences Division (ESD) Coral Bleaching Classifier dataset. The dataset includes human-annotated points indicating healthy and bleached coral, enabling classification for marine ecosystem monitoring.

Using latest YOLO architecture, deployment achieved real time classification speeds

  • Inference Speed~0.40 ms/image

  • Preprocessing Speed~0.17 ms/image

  • Postprocessing Speed~0.0003 ms/image

 

Sea Urchin Object Detection

 

Real time Fish Object Detection and Tracking