Coral-104 EfficientNet-B0 — V14
Overview
V14 is an EfficientNet-B0 image classification model developed for
automated identification of 104 coral species.
The model was trained using 5,172 images representing 104 species.
Model
- Architecture: EfficientNet-B0
- Framework: PyTorch
- Input: RGB image
- Input size: 224 × 224 pixels
- Number of classes: 104
- Best epoch: 30
- Random seed: 42
Performance
Independent test set:
- Test images: 726
- Correct predictions: 711
- Incorrect predictions: 15
- Accuracy: 97.93%
- 95% bootstrap CI: 96.83–98.90%
- Macro Precision: 98.26%
- Macro Recall: 97.93%
- Macro F1: 97.71%
- Top-3 accuracy: 98.21%
- Top-5 accuracy: 98.35%
Genus-level performance:
- Genus accuracy: 99.17%
- Genus Macro F1: 99.24%
Calibration
Temperature scaling was optimized using the validation set only.
Optimal temperature:
1.002572
The independent test set was not used to optimize the temperature.
Intended use
The model is intended as a research and educational tool for automated
coral image classification and decision support.
It should not be considered a replacement for expert taxonomic
identification.
Limitations
Performance may decrease for images that differ substantially from the
training distribution, including different geographic regions,
underwater imaging conditions, lighting, camera systems, coral growth
forms, partial colonies, damaged specimens, or previously unseen taxa.
Reproducibility
The model package contains:
- model weights
- class mappings
- preprocessing specification
- calibration information
- inference code
- configuration metadata
The independent test set was not used for model training or model
selection.
Citation
Please cite the associated peer-reviewed publication once available.
Version
V14
