akridge/noaa-esd-coral-bleaching-vit-classifier-v1

NOAA ESD Coral Bleaching ViT Based Patch Image Classifier

📝 Model Overview

This model was trained to classify coral bleaching conditions using the Vision Transformer (ViT) 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.

• Model Architecture: Google/ViT Base Patch16 224
• Task: Coral Bleaching Image Classification
• Classes:
  • CORAL: Healthy coral
  • CORAL_BL: Bleached coral
    

📊 Model Weights

• Download the TorchScript model here
• Download the ONNX model here
• Access the base model folder here

📅 Dataset & Annotations

• Dataset: NOAA ESD Coral Bleaching Classifier Dataset
• Annotation Method:
  • Points annotated by human experts using both randomly generated and targeted sampling methods.

Split	Images	Description
Training	7,292	Used for model training.
Validation	1,562	Used for model hyperparameter tuning and early stopping.
Test	1,565	Used for final model evaluation.

📚 Training Configuration

• Dataset: NOAA ESD Coral Bleaching Classifier Dataset
• Training/Validation Split: 70% training, 15% validation, 15% testing
• Epochs: 100
• Batch Size: 16
• Learning Rate: 3e-4
• Image Size: 224x224 (consistent with ViT input requirements)

📈 Results and Metrics

The model was evaluated using a withheld test set. The predictions were compared against human-labeled points for validation.

📄 Classification Report:

	precision	recall	f1-score
CORAL	0.86	0.91	0.88
CORAL_BL	0.84	0.75	0.79
accuracy			0.85
macro avg	0.85	0.83	0.84
weighted avg	0.85	0.85	0.85

🚀 How to Use the Model

🔗 Load with Transformers

# Load model directly
from transformers import AutoImageProcessor, AutoModelForImageClassification

processor = AutoImageProcessor.from_pretrained("akridge/noaa-esd-coral-bleaching-vit-classifier-v1")
model = AutoModelForImageClassification.from_pretrained("akridge/noaa-esd-coral-bleaching-vit-classifier-v1")

import torch
from transformers import ViTForImageClassification, AutoImageProcessor
from PIL import Image

# ✅ Load the model and processor
model = ViTForImageClassification.from_pretrained("akridge/noaa-esd-coral-bleaching-vit-classifier-v1")
processor = AutoImageProcessor.from_pretrained("akridge/noaa-esd-coral-bleaching-vit-classifier-v1")

# ✅ Load and process image
image = Image.open("your_image.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")

# ✅ Perform inference
with torch.no_grad():
    outputs = model(**inputs)
    prediction = outputs.logits.argmax(-1).item()

id2label = model.config.id2label
print(f"Prediction: {prediction} ({id2label[prediction]})")

🔗 Use TorchScript

import torch

# ✅ Load the TorchScript model
scripted_model = torch.jit.load("noaa-esd-coral-bleaching-vit-classifier-v1.pt")
scripted_model.eval()

# ✅ Inference with TorchScript model
with torch.no_grad():
    scripted_output = scripted_model(inputs["pixel_values"])
    scripted_prediction = scripted_output.argmax(-1).item()

print(f"TorchScript Prediction: {id2label[scripted_prediction]}")

🔗 Use ONNX

import onnxruntime as ort

# ✅ Load the ONNX model
onnx_model = "noaa-esd-coral-bleaching-vit-classifier-v1.onnx"
ort_session = ort.InferenceSession(onnx_model)

# ✅ Prepare ONNX input
onnx_inputs = {"input": inputs["pixel_values"].numpy()}

# ✅ Run inference with ONNX
onnx_outputs = ort_session.run(None, onnx_inputs)
onnx_prediction = onnx_outputs[0].argmax(axis=1)[0]

print(f"ONNX Prediction: {id2label[onnx_prediction]}")

Intended Use

• Monitoring coral reef health through automated image classification.
• Scientific research in marine biology and ecosystem science.

Limitations

• The model was trained on the NOAA ESD dataset; it may not generalize to different regions or unrepresented coral species.
• Images with low resolution or poor lighting may lead to incorrect predictions.
• Vertical or flipped images should be processed with appropriate orientation adjustments.

Ethical Considerations

• Predictions should not replace expert human validation in critical conservation decisions.

Metadata / Citation

Citation:
Pacific Islands Fisheries Science Center (2025). Ecosystem Sciences Division (ESD);

Related Metadata:

• Benthic Cover Derived from Analysis of Benthic Images (2019)
• NOAA ESD Coral Bleaching Classifier Annotations Data Dictionary
• Developing a semi-automated CoralNet Bleaching Classifier: annotations and imagery from survey sites across the Hawaiian Archipelago between 2014 and 2019

Disclaimer

This repository is a scientific product and is not official communication of the National Oceanic and Atmospheric Administration, or the United States Department of Commerce. All NOAA project content is provided on an ‘as is’ basis and the user assumes responsibility for its use. Any claims against the Department of Commerce or Department of Commerce bureaus stemming from the use of this project will be governed by all applicable Federal law. Any reference to specific commercial products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply their endorsement, recommendation or favoring by the Department of Commerce. The Department of Commerce seal and logo, or the seal and logo of a DOC bureau, shall not be used in any manner to imply endorsement of any commercial product or activity by DOC or the United States Government.