16,324 expert-verified phytoplankton ROIs across 42 classes, imaged by a Planktoscope at the Santa Cruz Municipal Wharf. Labels are the image folder names (HF imagefolder).

The companion classifier patcdaniel/planktoscope-phytoplankton-classifier currently predicts the subset of classes with ≥50 images; the rarer classes are included here for completeness and future model versions.

Loading

from datasets import load_dataset
ds = load_dataset("patcdaniel/planktoscope-phytoplankton")

How it was built

ROIs were grouped by DINOv2 deep-feature clustering (UMAP + HDBSCAN), reviewed by an expert, and rare classes were expanded by embedding-similarity mining against the unlabeled sample pool. review/ (unverified) images are excluded.

Class counts

class	images
Chaetoceros	3758
Pseudo-nitzschia	2259
Akashiwo	1799
Cerataulina_Guinardia_Dactyliosen	1629
Phaeocystis	1068
Tintinnid	948
Detritus	934
Lioloma	658
Rhizosolenia	614
Margalefidinium	517
Asterionella	409
Thalassionema	368
Tiarina	217
Tripos	165
Zooplankton	130
Pennate	105
Eucampia	78
Polykrikos	76
Dinophysis	69
Pleurosigma	54
Centric	51
Ditylum	49
Protoperidinium	45
Coscinodiscus	42
Strombidium_conicum	33
Ciliate_morpho_1	32
Dinoflagellate-Unknown	29
Chaetoceros_socialis	25
Thalassiosira	21
Ciliate	19
Scrippsiella	17
Mesodinium	17
Noctiluca	15
Prorocentrum_micans	14
Odontella	14
Cyst	12
Ditctyocha	10
Coccolith	8
Alexandrium	6
Hemialus	5
Gyrodinium	3
Oxyphysis	2

Notes

Counts are object counts, not biovolume.
Class imbalance is large; stratify when splitting.

License

cc-by-4.0

Downloads last month: 83

Total file size:

8.35 MB