FIESTA Scattering Bio — scattering transforms complement a plankton CNN

Headline change: the repository is now built around a stacked CNN + scattering meta-classifier instead of the prior 10-class classification demo.

Result (held-out, Decrop et al. 2025's test.txt)

| Method | Top-1 | Top-5 | Rare-class mean recall | |---|---:|---:|---:| | CNN alone (Decrop 2025) | 86.34 % | 98.70 % | 47.7 % | | Scattering + LR alone | 26.93 % | 60.43 % | 43.0 % | | 50/50 probability ensemble | 86.28 % | 94.82 % | 50.3 % | | Stacked LR (val-trained) | 85.62 % | 95.39 % | 56.1 % | | Oracle ceiling | 87.68 % | — | 64.6 % |

• 33 of 95 classes improve by > 1 pp under stacking; only 16 degrade.
• +8.4 pp rare-class recall — captures about half the oracle ceiling.
• −0.72 pp overall top-1 — the bounded trade-off.

Full per-class numbers in `results/stacking_val_trained_results.json`.

What changed

• Three-step pipeline (`01_scattering_features.py`, `02_cnn_predictions.py`, `03_stacking.py`) replaces the old single-notebook 10-class experiment.
• New companion repo fiesta-decrop-reproduction produces the CNN predictions on Decrop's val + test splits — reused here as the CNN baseline.
• Dockerfile rewritten: apt installs p7zip-full/git/libgl1/libglib2.0-0, pip install uses `--extra-index-url` for the torch CPU wheel (fixes the v0.2.0 build failure caused by `--index-url` overriding PyPI for non-torch packages).
• `environment.yml` expanded; new optional `environment-cnn.yml` for in-repo CNN inference.
• CITATION.cff and codemeta.json now cite Decrop et al. 2025 as the baseline paper (DOI 10.3389/fmars.2025.1699781).

Why this framing

CNNs on class-imbalanced microscopy datasets systematically underperform on rare taxa. Scattering coefficients are deterministic wavelet statistics with no learned parameters and no frequency bias. The two predictors err on different samples — scattering alone is uniquely correct on 1.3 % of test images, heavily enriched in rare classes — and a small linear meta-classifier trained on Decrop's held-out val split captures most of this complementarity.

This is the one FIESTA-OSCARS repository demonstrating scattering as a complementary layer on top of an existing deep model, alongside three gap-filling repos (astro, SST, SST-WGS84) and the CNN reproduction repo.

Breaking changes

• `01_plankton_classification.py`, `02_cnn_baseline.py`, `03_reproduce_decrop.py` removed. Git history preserves them.
• `experiments/` folder removed (scripts consolidated to top level).
• Old `results/plankton_classification_results.json` removed. The canonical artefact is now `results/stacking_val_trained_results.json`.

🤖 Generated with Claude Code