Reproduction of Decrop et al. 2025 — phytoplankton CNN classifier

First release of the repository. Contains two notebooks that together constitute a FAIR computational reproduction of Decrop et al. 2025 plus a downstream artefact used by fiesta-scattering-bio.

Notebooks

| Notebook | Input | Output | Purpose | |---|---|---|---| | `01_reproduce_decrop.py` | `test.txt` (33,718 images) | `results/reproduce_decrop_results.json` + `reproduce_decrop_predictions.npz` | Reproduce the five published metrics | | `02_cnn_val_predictions.py` | `val.txt` (33,829 images) | `results/cnn_predictions_val.npz` | Provide CNN predictions on the held-out val split for downstream stacking in fiesta-scattering-bio |

Both notebooks share an identical pipeline — same pretrained EfficientNetV2-B0 weights, same preprocessing, same 10-crop test-time augmentation. Only the split differs.

Reproduction outcome on `test.txt`

Every reported metric matches the paper to within 0.003 percentage points:

| Metric | Paper | This reproduction | Delta (pp) | |---|---:|---:|---:| | Top-1 accuracy | 86.34 % | 86.3426 % | +0.003 | | Top-5 accuracy | 98.76 % | 98.7633 % | +0.003 | | Micro F1 | 86.34 % | 86.3426 % | +0.003 | | Macro F1 | 78.76 % | 78.7583 % | −0.002 | | Weighted F1 | 86.25 % | 86.2470 % | −0.003 |

Inference took 27 minutes on a single M1 Pro CPU pool. All inputs are openly archived: dataset (Zenodo 10554845), weights (Zenodo 15269453), code (planktonclas, Apache 2.0). A third party can rerun this end-to-end with zero guessing.

What this release enables

• Versioned Zenodo DOI for this reproduction, citable alongside Decrop et al. 2025.
• Docker image on GHCR: `ghcr.io/annefou/fiesta-decrop-reproduction:0.2.0` — runs the whole pipeline in one `docker run`.
• CNN prediction artefacts (`cnn_predictions_test.npz`, `cnn_predictions_val.npz`) consumable by the stacked CNN + scattering experiment in fiesta-scattering-bio v0.3.0.

FIESTA-OSCARS context

Part of the five-repo FIESTA-OSCARS project. This repo plays two roles:

1. Standalone FAIR reproduction of a published CNN classifier.
2. Upstream dependency for the biodiversity scattering-stacking experiment.

🤖 Generated with Claude Code