CSIRO Patch-based 9-class Morphology Model (Australia)

Input

The model processes benthic quadrat images and divides each image into fixed-size patches for classification. It was developed primarily for imagery collected from Australian seagrass habitats, including Shark Bay, Queensland, Lizard Island, Torres Strait and Western Australian Posidonia habitats.

The model may process imagery from other locations, but its accuracy may decrease when the species composition, environmental conditions, camera configuration or image-acquisition methods differ substantially from those represented in the training dataset.

Classes

This model classifies image patches into the following seagrass morphology categories:

• Oval Seagrass — Ho, Hd, Hb, Hm
• Cylindrical Seagrass — Si
• Stemmy Seagrass — Tc, Ag, Aa
• Ferny Seagrass — Hs
• Strappy Medium/Thin Seagrass — Cs, Th, Ps, Pc
• Strappy Narrow/Hair-Thin Seagrass — Hu, Cr, Zc, Hp
• Strappy Thick Seagrass — Ea, Pa
• Unknown or Mixed Seagrass
• Others/Background

Note: The italicised codes correspond to the seagrass species represented within each morphology group. Model predictions should normally be interpreted at the morphology level rather than as definitive species-level identifications.
For full species names and descriptions, refer to the Seagrass-Watch Species ID Guide.

Output

For each input image, the model produces a patch-level classification map that approximates a coarse segmentation mask.

The patch predictions are aggregated to estimate:

• total seagrass percentage cover;
• percentage cover for each seagrass morphology class; and
• the approximate spatial distribution of seagrass morphologies within the image.

The outputs are model-derived estimates and should not be interpreted as pixel-accurate segmentation or direct field measurements.

Training Data

The Australian model was trained using 60,477 labelled patches from 6,462 source images. The training imagery includes data from Shark Bay, Queensland, Lizard Island, Torres Strait-related surveys and Western Australian Posidonia datasets.

The distribution of the training patches is:

• Oval Seagrass: 14,577
• Others/Background: 13,973
• Ferny Seagrass: 9,851
• Strappy Narrow/Hair-Thin Seagrass: 8,187
• Stemmy Seagrass: 4,971
• Strappy Medium/Thin Seagrass: 3,878
• Unknown or Mixed Seagrass: 3,204
• Strappy Thick Seagrass: 1,202
• Cylindrical Seagrass: 634

Annotation Method

Each image represents a 50x50 cm quadrat captured by a camera positioned at a fixed distance. The images are resized to 3008x3008 pixels, achieving a resolution of approximately 60 pixels per cm.

Patch Size

280x280 pixels: Corresponds to a 5x5 cm area.

Annotated Patches on Reefcloud

Due to software limitations, annotators are required to label 20 patches per image slice, arranged in a 5x4 grid. This setup allows for a maximum of 1024 patches (32x32 grid) per image.

Patch Labeling Methodology

Single Seagrass Species Present: Assign the specific seagrass class label to the patch, regardless of the percentage of coverage within the patch.​

Multiple Seagrass Species Present:

• If one species is clearly dominant, label the patch with the dominant species’ class.
• If no species is dominant, use a generic seagrass class label.​
• Seagrass vs. Benthos Classes: Seagrass classifications take precedence over benthic classes. If no seagrass is visible, label the patch with the dominant benthic class.

Contextual Considerations

Annotators focus on the content within the patch while also considering the broader context of the entire image. This approach ensures that local features are interpreted in alignment with the global scene, leading to more accurate and meaningful annotations.

Limitations

• The model was developed for Australian benthic quadrat imagery and may be less reliable when applied to regions, habitats, species communities or imaging conditions that are not represented in the training dataset.
  Predictions are produced at the patch level and therefore provide coarse spatial information rather than precise boundaries around individual seagrass shoots or blades.
• Where multiple seagrass morphologies occur within the same patch, the model generally predicts the dominant or most visually interpretable morphology. Minority morphologies may consequently be under-represented.
• The model identifies morphology groups rather than providing definitive species-level classifications. Visually similar species within the same morphology group cannot always be distinguished reliably.
• Turbidity, glare, shadows, colour imbalance, motion blur, inconsistent camera height, low seagrass cover and substrate occlusion may reduce prediction reliability.
• Some morphology classes contain substantially fewer training samples than others. Predictions for rare classes, particularly cylindrical and strappy thick seagrass, should therefore be interpreted cautiously.
• All non-seagrass substrates are grouped into a single Others/Background class. The model is not intended to provide detailed classification of coral, sand, rubble, mud or other non-seagrass benthic habitats.