README.md

metadata

annotations_creators:
  - expert-generated
language:
  - en
license: cc-by-4.0
pretty_name: IFCB Plankton Labeled (Cluster-Sorted)
task_categories:
  - image-classification
tags:
  - plankton
  - microscopy
  - ifcb
  - clustering
  - manual-sorting
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': Akashiwo
            '1': Asterionellopsis
            '2': Centric
            '3': Ceratium
            '4': Chaetoceros
            '5': Ciliates
            '6': Detonula
            '7': Diatom_aggregate
            '8': Dictyocha
            '9': Dinophysis
            '10': Ditylum
            '11': Eucampia
            '12': Guinardia
            '13': Gymnodinium
            '14': Gyrodinium
            '15': Heterocapsa
            '16': Lauderia_Hemiaulus
            '17': Leptocylindrus
            '18': NanoPlankton
            '19': Pennate
            '20': Phaeocystis
            '21': Prorocentrum
            '22': Protoperidinium
            '23': Psuedo-nitzschia
            '24': Rhizosolenia_Proboscia
            '25': Skeletonema
            '26': Stephanopyxis
            '27': Thalassionema
            '28': Thalassiosira
            '29': Tintinnid
            '30': Unknown_detritus
            '31': Unknown_full_image
            '32': Unknown_phyto
    - name: group
      dtype: string
    - name: filename
      dtype: string
  splits:
    - name: train
      num_bytes: 286871203.36
      num_examples: 7895
  download_size: 291443502
  dataset_size: 286871203.36

IFCB Plankton Labeled (Cluster-Sorted)

This dataset contains labeled images of phytoplankton collected with the Planktivore Imaging System. Images were preprocessed with a zero-padding and resized to the standard size used for ViT_b_16

The dataset was originally constructed by clustering unlabeled ROI images using deep features from a ViT model.
Clusters were then saved locally and manually curated into taxonomic labels and higher-order groups.

Dataset Summary

• Modality: Images (PNG)
• Source: Planktivore ROI captures
• Curation process:
  1. Extracted deep features with a ViT backbone.
  2. Applied clustering (UMAP + HDBSCAN) to group morphologically similar images.
  3. Exported clusters to local folders.
  4. Manually reviewed and sorted each cluster into taxonomic categories (label) and broader groups (group).

Columns

• image: The plankton ROI image.
• label: Fine-grained label (taxon).
• group: Higher-order grouping (e.g. diatoms, dinoflagellates, ciliates).

Example

from datasets import load_dataset

ds = load_dataset("patcdaniel/synchro-April2025-cluster-labeled-highMag")
sample = ds["train"][0]
sample["image"].show()
print("Label:", ds["train"].features["label"].int2str(sample["label"]))
print("Group:", ds["train"].features["group"].int2str(sample["group"]))