Mussel-RGB: Mussel Segmentation Model for RGB Drone Imagery

Model Type: ONNX Semantic Segmentation
Application: Mussel detection in high-resolution RGB aerial imagery
Input: 3-band RGB imagery (Red, Green, Blue)
Output: Binary segmentation mask (mussel vs. non-mussel)

Model Description

The Mussel-RGB model is a deep learning semantic segmentation model specifically trained for detecting mussels in RGB drone imagery. This model enables precise mussel bed mapping and monitoring for marine habitat assessment, aquaculture applications, and intertidal ecology research.

Key Features:

Specialized for mussel detection in RGB imagery
Optimized for standard RGB imagery from drones
ImageNet-pretrained normalization statistics
Efficient ONNX format for cross-platform deployment
Designed for intertidal and subtidal mussel monitoring

Model Details

Version: 20250711
Input Channels: 3 (RGB)
Input Size: Dynamic tiling (recommended: 2048x2048 tiles)
Normalization: Standard (ImageNet statistics)
Output: Binary segmentation (0: background, 1: mussels)
Format: ONNX

Normalization Parameters

The model expects input images to be normalized using ImageNet statistics:

{
  "mean": [0.485, 0.456, 0.406],
  "std": [0.229, 0.224, 0.225],
  "max_pixel_value": 255.0
}

Usage

1. Using kelp-o-matic CLI (recommended)

For command-line usage:

# Install kelp-o-matic
pip install git+https://github.com/HakaiInstitute/kelp-o-matic@dev

# List available models
kom list-models

# Run segmentation on mussel RGB drone imagery
kom segment \
    --model mussel-rgb \
    --input /path/to/mussel_rgb_image.tif \
    --output /path/to/mussel_segmentation.tif \
    --batch-size 8 \
    --crop-size 2048 \
    --blur-kernel 5 \
    --morph-kernel 3

# Use specific model version
kom segment \
    --model mussel-rgb \
    --version 20250711 \
    --input image.tif \
    --output result.tif

# For high-resolution mussel surveys
kom segment \
    --model mussel-rgb \
    --input high_res_mussel_survey.tif \
    --output result.tif \
    --batch-size 4 \
    --crop-size 1024

2. Using kelp-o-matic Python API

The easiest way to use this model is through the kelp-o-matic package:

from kelp_o_matic import model_registry

# Load the model (automatically downloads if needed)
model = model_registry["mussel-rgb"]

# Process a large mussel survey image with automatic tiling
model.process(
    input_path="path/to/your/mussel_rgb_image.tif",
    output_path="path/to/output/mussel_segmentation.tif",
    batch_size=8,  # Higher batch size for RGB
    crop_size=2048,
    blur_kernel_size=5,  # Post-processing median blur
    morph_kernel_size=3,  # Morphological operations
)

# For more control, use the predict method directly
import rasterio
import numpy as np

with rasterio.open("mussel_survey_image.tif") as src:
    # Read a 2048x2048 tile (3 bands: RGB)
    tile = src.read(window=((0, 2048), (0, 2048)))  # Shape: (3, 2048, 2048)
    tile = np.transpose(tile, (1, 2, 0))  # Convert to HWC
    
    # Add batch dimension and predict
    batch = np.expand_dims(tile, axis=0)  # Shape: (1, 2048, 2048, 3)
    batch = np.transpose(batch, (0, 3, 1, 2))  # Convert to BCHW
    
    # Run inference (preprocessing handled automatically)
    predictions = model.predict(batch)
    
    # Post-process to get final segmentation
    segmentation = model.postprocess(predictions)
    # Result: 0=background, 1=mussels

3. Direct ONNX Runtime Usage

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from PIL import Image

# Download the model
model_path = hf_hub_download(repo_id="HakaiInstitute/mussel-rgb", filename="model.onnx")

# Load the model
session = ort.InferenceSession(model_path)

# ImageNet normalization parameters
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])

# Preprocess your RGB image
def preprocess(image):
    """
    Preprocess RGB image for model input
    image: numpy array of shape [height, width, 3] with pixel values 0-255
    """
    # Normalize to 0-1
    image = image.astype(np.float32) / 255.0
    
    # Apply ImageNet normalization
    image = (image - mean) / std
    
    # Reshape to model input format [batch, channels, height, width]
    image = np.transpose(image, (2, 0, 1))  # HWC to CHW
    image = np.expand_dims(image, axis=0)  # Add batch dimension
    
    return image

# Load and preprocess image
image = np.array(Image.open("mussel_bed_drone_image.jpg"))
preprocessed = preprocess(image)

# Run inference
input_name = session.get_inputs()[0].name
output = session.run(None, {input_name: preprocessed})

# Postprocess to get binary mask
logits = output[0]
prediction = np.argmax(logits, axis=1).squeeze(0).astype(np.uint8)
# Result: 0=background, 1=mussels

4. Using HuggingFace Hub Integration

from huggingface_hub import hf_hub_download
import onnxruntime as ort

# Download and load model
model_path = hf_hub_download(
    repo_id="HakaiInstitute/mussel-rgb",
    filename="model.onnx",
    cache_dir="./models"
)

session = ort.InferenceSession(model_path)
# ... continue with preprocessing and inference as above

Installation

For kelp-o-matic usage:

# Via pip
pip install git+https://github.com/HakaiInstitute/kelp-o-matic@dev

For direct ONNX usage:

pip install onnxruntime huggingface-hub numpy pillow
# For GPU support:
pip install onnxruntime-gpu

Input Requirements

Image Format: 3-band RGB raster (JPEG, PNG, GeoTIFF)
Band Order: Red, Green, Blue
Pixel Values: Standard 8-bit (0-255 range)
Spatial Resolution: Optimized for high-resolution drone imagery (cm-level)
Scene Type: Intertidal zones, shallow subtidal, aquaculture sites with mussel beds

Output Format

Type: Single-band raster with class labels
Values:
- 0: Background (water, rocks, other substrates, non-mussel organisms)
- 1: Mussels (typically blue mussels, Mytilus spp.)
Format: Matches input raster format and projection
Spatial Resolution: Same as input

Note: The model outputs class probabilities, but kelp-o-matic automatically applies argmax to convert these to discrete class labels.

Performance Notes

Dynamic Tile Size: Supports flexible tile sizes (recommended: 2048x2048 or 1024x1024)
Batch Size: Start with 4, adjust based on available GPU memory

Large Image Processing

For processing large geospatial images, the kelp-o-matic package handles:

Automatic Tiling: Splits large images into manageable tiles
Overlap Handling: Uses overlapping tiles to avoid edge artifacts
Memory Management: Processes tiles in batches to manage memory usage
Geospatial Metadata: Preserves coordinate reference system and geotransforms
Post-processing: Optional median filtering and morphological operations

Citation

If you use this model in your research, please cite:

@software{Denouden_Kelp-O-Matic,
  author = {Denouden, Taylor and Reshitnyk, Luba},
  doi = {10.5281/zenodo.7672166},
  title = {{Kelp-O-Matic}},
  url = {https://github.com/HakaiInstitute/kelp-o-matic}
}

License

MIT License - see the kelp-o-matic repository for details.

Related Resources

Documentation: kelp-o-matic.readthedocs.io
Source Code: github.com/HakaiInstitute/kelp-o-matic
Other Models: Check the Hakai Institute HuggingFace organization for additional marine segmentation models

Contact

For questions or issues:

Open an issue on the GitHub repository
Contact: Hakai Institute

HakaiInstitute
/

mussel-rgb