patcdaniel/planktoscope-phytoplankton-classifier

Planktoscope Phytoplankton Classifier

A 22-class phytoplankton image classifier for Planktoscope ROIs from the Santa Cruz Municipal Wharf timeseries.

Architecture

A frozen DINOv2 ViT-S/14 backbone (timm: vit_small_patch14_dinov2.lvd142m) produces a 384-d embedding, and a single linear layer (linear probe) maps it to 22 classes. Only the linear head is trained; the backbone is unchanged.

Preprocessing (must match exactly)

1. Pad to square with fill (255, 255, 255) (light, matching the ROI background) — preserves aspect ratio so chains are not distorted.
2. Resize to 224×224.
3. Normalize with ImageNet mean [0.485, 0.456, 0.406] / std [0.229, 0.224, 0.225].

Classes

Akashiwo, Asterionella, Centric, Cerataulina_Guinardia_Dactyliosen, Chaetoceros, Detritus, Dinophysis, Eucampia, Lioloma, Margalefidinium, Pennate, Phaeocystis, Pleurosigma, Polykrikos, Protoperidinium, Pseudo-nitzschia, Rhizosolenia, Thalassionema, Tiarina, Tintinnid, Tripos, Zooplankton

Usage

See inference_example.py (self-contained — does not require the training repo). Predictions below recommended_min_confidence (0.7) should be treated as Unassigned, important when running over raw samples that contain detritus and unfamiliar particles.

Per-class performance (held-out test)

class	support	precision	recall	f1
Dinophysis	10	1.0	1.0	1.0
Phaeocystis	160	0.964	1.0	0.982
Thalassionema	55	0.982	0.982	0.982
Tripos	25	0.962	1.0	0.98
Tintinnid	142	0.959	0.993	0.976
Akashiwo	270	0.989	0.963	0.976
Zooplankton	19	1.0	0.947	0.973
Chaetoceros	564	0.97	0.968	0.969
Detritus	140	0.985	0.95	0.967
Cerataulina_Guinardia_Dactyliosen	244	0.963	0.967	0.965
Margalefidinium	78	0.928	0.987	0.957
Tiarina	33	0.917	1.0	0.957
Asterionella	61	0.951	0.951	0.951
Centric	8	1.0	0.875	0.933
Pseudo-nitzschia	339	0.944	0.903	0.923
Lioloma	99	0.873	0.97	0.919
Polykrikos	11	0.846	1.0	0.917
Protoperidinium	10	0.833	1.0	0.909
Eucampia	12	0.846	0.917	0.88
Rhizosolenia	92	0.907	0.848	0.876
Pennate	16	0.786	0.688	0.733
Pleurosigma	8	0.667	0.75	0.706

Training data

patcdaniel/planktoscope-phytoplankton — expert-verified ROIs curated via deep-feature clustering + DINOv2 embedding-similarity mining. This model predicts the 22 classes with ≥50 training images; the dataset additionally includes rarer curated classes not yet covered by the model.

Caveats

• Abundances derived from this model are counts of particles, not biovolume — a single cell and a long chain each count once.
• A few small classes remain weak (low support); see the per-class table.

License

cc-by-4.0