Data, code, model weights, and validation materials supporting the manuscript “Automated Shadowgraph Imaging and Deep Learning for Sub-millimetre Plankton Imaging and Rare-Target Screening” (Ecological Informatics, ECOINF-D-26-00458, revision 2). This version (REV2) supersedes the pre-REV1 deposit: all pre-REV1 objects are retained unchanged, and a new rev2ᵥalidation/ directory adds the detection-stage characterisation and blind multi-rater validation produced for the revision. Download and extract ECOINF-D-26-00458dataₐndcodeREV2. zip; a top-level README. md and a file-level metadata. csv (with SHA-256 checksums for every file) document the contents, and rev2ᵥalidation/README. md provides a figure/table → file reproduction crosswalk and data dictionary. Contents include: Curated training set: 41, 487 Lepeophtheirus salmonis copepodid shadowgraph images (positive class) and 44, 601 ambient-plankton / hard-negative images (negative class). Trained ConvNeXt binary classifier (architecture and weights). Full C++/Python acquisition, segmentation, inference, metadata-enrichment and compression pipeline. A random 100, 001-ROI Austevoll June 2025 field subset, and all 192 high-probability candidate ROIs (64 June + 128 October). A 36, 179-ROI Flødevigen catalogue with the PaCMAP atlas builder and the iterative cosine-similarity annotation tool. REV2 additions: the blind four-class annotation tool and the detector gate-recall harness; per-annotator votes, provenance and derived statistics; the degraded positive-control stimulus sets; the source data, figures and analysis scripts behind Figures S1–S7 and Tables S1–S8; and the annotation decision codebook. Licences: data, ROI archives, images, figures, metadata and model weights are released under CC-BY-4. 0; source code under the MIT licence. Annotator identities in the blind validation are replaced throughout by expertise roles; the role-to-identity key is held privately by the corresponding author.
Mats Huserbräten (Thu,) studied this question.