sosiklab/NES-plankton-classifier-2022-dataset

NES Plankton Classifier 2022 Training Data

Image classification dataset of plankton species imaged by Imaging FlowCytobot (IFCB) on the Northeast US Shelf. Used to train the NES Plankton Classifier 2022.

Dataset Summary

• 155 classes of plankton and non-plankton ROIs (including detritus, bubbles, fibers, etc.)
• 97,026 labeled images across train and validation splits
• Images are grayscale PNGs extracted from IFCB sample files

Usage

Here's the easiest way to access parquet data.

from datasets import load_dataset

ds = load_dataset("sbatchelder/NES-plankton-classifier-2022-dataset")

example = ds["train"][0]
example["image"]        # PIL.Image.Image
example["label"]        # int class index
example["classname"]    # human-readable class name
example["ifcb_roi_pid"] # IFCB ROI persistent identifier

# Map int label -> name via ClassLabel metadata
ds["train"].features["label"].int2str(example["label"])

Exporting Parquet to Filesystem Images and Tabular CSV

Sometimes it's easier to work directly in the filesystem and simple dataformats. Use the following code to convert this repo's dataset into png images and CSV tabular data.

from pathlib import Path
import csv
from datasets import load_dataset


# Export Functions #

def save_images(ds, output_dir, split_subfolders=False, classname_subfolders=False, clobber=False):
    paths = {}
    for split_name, dset in ds.items():
        for ex in dset:

            path = Path(output_dir)
            if split_subfolders:
                path = path / split_name
            if classname_subfolders:
                path = path / ex["classname"]
            path.mkdir(parents=True, exist_ok=True)

            path = path / f'{ex["ifcb_roi_pid"]}.png'
            if clobber or not path.exists():
                ex["image"].save(path)

            paths[ex["ifcb_roi_pid"]] = str(path)

    return paths


def save_table_csv(ds, csv_path):
    csv_path = Path(csv_path)
    csv_path.parent.mkdir(parents=True, exist_ok=True)

    with csv_path.open("w", newline="") as f:
        writer = csv.writer(f)
        writer.writerow(["ifcb_roi_pid", "dataset", "class_label", "class_name"])

        for split, dset in ds.items():
            for ex in dset:
                writer.writerow([
                    ex["ifcb_roi_pid"],
                    split,
                    ex["label"],
                    ex["classname"],
                ])

    return str(csv_path)


# Run Dataset png and csv Export functions #

ds = load_dataset("sbatchelder/NES-plankton-classifier-2022-dataset")

image_paths = save_images(
    ds,
    "image_data/NES-plankton-classifier-2022-dataset",
    split_subfolders=True,
    classname_subfolders=True,
)

csv_path = save_table_csv(ds, "NES-plankton-classifier-2022-dataset.csv")

Columns

Column	Type	Description
image	Image	Grayscale ROI image from IFCB
label	ClassLabel	Integer class index
classname	string	Human-readable class name (same as label decoded)
ifcb_roi_pid	string	IFCB ROI persistent identifier

Splits

Split	Examples
train	77,617
validation	19,409

License

MIT

Citation

If you use this dataset in your research, please cite the Woods Hole Oceanographic Institution and the Sosik Lab. A formal citation will be added here upon publication.

Acknowledgments

Developed at the Woods Hole Oceanographic Institution. Data collected by the WHOI IFCB Lab on Northeast U.S. Shelf cruises.