@prefix this: . @prefix sub: . @prefix np: . @prefix rdf: . @prefix prov: . @prefix npx: . @prefix dc: . @prefix xsd: . sub:Head { this: a np:Nanopublication; np:hasAssertion sub:assertion; np:hasProvenance sub:provenance; np:hasPublicationInfo sub:pubinfo . } sub:assertion { sub:decrop-2025-cnn-scattering-stacking-study a , ; "CNN + scattering stacking on Decrop 2025's plankton splits"; , , , , , ; "This is an extension, not a reproduction. Differences from Decrop et al. 2025: (1) we add a stacking meta-classifier on top of the CNN's softmax outputs; the original paper reports CNN-alone metrics. (2) We compute scattering features as additional input — the paper does not use scattering. (3) Reported metrics include mean rare-class recall (computed on the 13 classes with train-set size <200), which the paper does not report; aggregate top-1, top-5, micro/macro/weighted F1 are all from the same test.txt evaluation pipeline. Differences from Delouis et al. 2022 (FOSCAT method paper): we use FOSCAT's 2D scattering operator (scat_cov.funct) on flat RGB images, not the spherical HEALPix variant the paper develops; same library, different operator. As with chains #1–#4, FOSCAT was used via the annefou/FOSCAT@v0.1.0-cpu fork; that CPU patch has since been merged upstream and is included in foscat>=2026.4.1 on PyPI."; ; """Three-stage pipeline. Stage 1 (01_scattering_features.py): compute multi-scale scattering features on RGB phytoplankton images via FOSCAT scat_cov.funct(NORIENT=8, KERNELSZ=3) at 64×64 pixels, producing 246-dimensional feature vectors per image. Features are computed on (a) a balanced training subset, (b) Decrop's released val.txt (33,829 images), and (c) Decrop's released test.txt (33,718 images). Stage 2 (02_cnn_predictions.py): obtain CNN softmax probabilities for the same val and test images by running Decrop's pretrained EfficientNetV2-B0 via the authors' planktonclas package with 10-crop test-time augmentation. These predictions can equivalently be sourced from the upstream fiesta-decrop-reproduction repository (Zenodo 10.5281/zenodo.19701133), which is the FAIR-archived form of the same computation. Stage 3 (03_stacking.py): fit a class-weighted scikit-learn LogisticRegression on the 246-dim scattering features (StandardScaler + class_weight='balanced') to obtain scattering-derived softmax probabilities; concatenate with CNN softmax probabilities to form 190-dim meta-features; train a second class-weighted LogisticRegression on Decrop's val split as the stacking meta-classifier; evaluate on Decrop's test split. Compute top-1, top-5, and per-class recall against the integer labels in test.txt. Mean rare-class recall is averaged over the 13 'rare' classes defined as those with fewer than 200 training-set examples in Decrop's train.txt. Comparator results: CNN alone, scattering+LR alone, naive 50/50 probability ensemble, and a hard-switch oracle ceiling (perfect choice between CNN and scattering predictions per image) for context."""; "This study tests whether multi-scale scattering features computed independently of the CNN can be combined with the CNN's softmax probabilities via a class-weighted stacking meta-classifier to lift mean rare-class recall on Decrop et al. 2025's exact 95-class phytoplankton classification benchmark, at acceptable cost to aggregate top-1 accuracy. The chain reuses Decrop's released dataset, pretrained CNN weights, and train.txt/val.txt/test.txt splits, asking whether the rare-species class-imbalance limitation acknowledged in the paper can be partially addressed by complementary feature representations."; . } sub:provenance { sub:assertion prov:wasAttributedTo . } sub:pubinfo { "Anne Fouilloux" . this: dc:created "2026-04-26T19:36:55.873Z"^^xsd:dateTime; dc:creator ; dc:license ; npx:introduces sub:decrop-2025-cnn-scattering-stacking-study; npx:wasCreatedAt ; "NP created using Declaring a replication study design according to FORRT"; . sub:sig npx:hasAlgorithm "RSA"; npx:hasPublicKey "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"; npx:hasSignature "PIA4Um5UQH4xMFLp/s8iFK5Gg2IVSBL3YbYT7FczqTbUBYGtvebzdgYhqiKbIhsuhEVLbRwLmbvfBOOgTImcti7GEwkq5mevnVvSczPHURTo2+zTydhyNdj02O1XKurHX7jyKWo0ctwCK0vPnDzBc28Tq0CJrMH4KrEkxJBhOiL8htULWwURyzDvjVqvEDrYrqSP0YR0w4NmEtc6nz6XEt695BJqU4q420F2A+s2aSY6jgRAF4YFNNjPfcTg4e//IWpHgkKiNhSeEGOtxepPJxu1TGfEeApT/LjQjNvaEyMjhDuzwXGwDTG3CI1y2JVanpwhWa9sg1K6mfOWU3Bl1BMOKSfhxJ2a4IovHAB7UeI1xmOYShn9MftfJpK2vD+sCXjfT2MWj0IdcgfHCX1ytEgPgHfMHlVLPe3/I77y5ZjsURwv0OmvbN49WPbysmp3/CJpgYYoN0aDg0XKe4HWauHJEMd4lzVrOdkHlt+1s/kDQlJuTpho6gpSTgQWFERqHaWVC85V1oRAjjdxwMinPm3sG4YxgT9Njd71u7NA5LHqPje24Zk9k+wttDmtLHJF+1xlZGIsADjhWwS7LFBuEl6NV1Sno/xfIQcfIzXdXHbCC7LhCXostZNjLNRQw0xaYuEHRj0KAqGKYh8yVSYhOnlptXkwF4jPmKwuJqCtPbA="; npx:hasSignatureTarget this:; npx:signedBy . }