HakaiInstitute/mussel-seg-1024-1024

MusselSeg: Semantic Segmentation for Rocky Intertidal Mussel Habitat

Dataset description

MusselSeg is a large-scale dataset for semantic segmentation of mussel habitat using high resolution drone imagery. It covers coastal mussel habitat located on the central coast of British Columbia, Canada, as well as areas in California, USA and provides pixel-wise annotation for mussel beds.

• —Source: Imagery collected by the Hakai Institute and University of California Santa Cruz
• —Geographic Coverage: ~3500m²
• —Time Period: 2021-2023

Task description

The dataset is designed for semantic segmentation of mussel habitat in aerial imagery. The task involves assigning each pixel in the image to either the "mussel" class or "background" (i.e. not mussels) class.

Usage

Download and iterate

Install the HuggingFace datasets library (instructions)

from datasets import load_dataset

train_dataset = load_dataset("HakaiInstitute/mussel-seg-1024-1024", split="train")
val_dataset = load_dataset("HakaiInstitute/mussel-seg-1024-1024", split="validation")
test_dataset  = load_dataset("HakaiInstitute/mussel-seg-1024-1024", split="test")

for sample in train_dataset:
    x = sample["img.tif"]
    y = sample["seg.tif"]
    # x and y are `PIL.Image` instances, ready to feed into a training loop, PyTorch dataloader, etc.

    # ...

Streaming from HuggingFace

This data is released as a WebDatasets, which makes it possible to use the data without downloading it in advance. For instructions on how to do this, please see WebDataset

Data characteristics

• —Image Format: GeoTiff
• —Resolution: mean=0.45cm, stdev=0.20cm
• —Tile Size: 1024x1024 pixels with 50% overlap
• —Number of Tiles: 9972 image and label pairs
• —Total Dataset Size: 42G

Annotation details

• —Method: Manual heads-up digitizing with manual verification
• —Format: Pixel-wise labels stored as separate mask images
• —Labelling Convention: Each pixel assigned a single class label

Class distribution

Split information

Train and Validation split tiles all contain at least 1 pixel in each class. For the Test split, some tiles are entirely the background class. If ignoring the test split tiles which contain only background pixels, the split percentages instead become 70/17/13 for the train/validation/test splits, respectively.

Splits are created such that tiles from the same source orthomosaic image are not divided across different splits. That is, all tiles from the same drone flight are present only in a single split.

Preprocessing

1. 1.Orthorectification applied to raw imagery
2. 2.Tiles extracted with 50% overlap
3. 3.Tiles with no mussels present eliminated for the Train and Validation splits

Licensing information

This dataset is released under the Creative Commons Attribution 4.0 License (CC BY 4.0).

Ethical considerations

• —No identifiable individuals are present in imagery
• —Minimized impact on wildlife and sensitive habitats
• —Engaged with local First Nations in planning aerial surveys

Citation information

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

@misc{denouden2024musselseg,
  author = {Denouden, Taylor and McInnes, William and Ammann, Karah and Fletcher, Nathaniel},
  title = {MusselSeg: Semantic Segmentation for Rocky Intertidal Mussel Habitat},
  month = July,
  year = 2024,
  doi = { 10.57967/hf/2760 },
  publisher = {Hakai Institute {\tt data@hakai.org}},
  howpublished = {\url{https://huggingface.co/datasets/HakaiInstitute/mussel-seg-1024-1024}}
}

Known limitations

• —Limited seasonal variation due to imagery being captured primarily in summer months
• —Imagery only covers areas with known mussel beds
• —No examples of mussel beds near urban or built-up environments
• —Labelling errors may be present in areas with shadows, where it is difficult to distinguish mussels beds