Nordic country-held-out direct-pollen predictions¶
This is a prior-exposed exploratory prediction of a closed 16-label, exact-name Neotoma pollen subcomposition. It predicts observed pollen composition; it is not an LRA reconstruction, measured vegetation estimate, or confirmatory test.
The entire target country is omitted from each prediction's training set. Denmark, Finland and Norway contribute the source-backed physical site groups counted below; Sweden contributes the three previously admitted Neotoma training sites. Targets use selected sample-owned Neotoma ages and site coordinates only. Prediction files are saved and hashed before the scoring command reads held-out count profiles.
The fixed methods are training mean, nearest site, inverse-distance-squared, and the existing CLR/WGS84 RBF-ridge (half-count smoothing on training cells, median positive training-distance scale, ridge 0.01). They are comparisons, not a Nordic-tuned winner search. No target-country outcome is used to choose taxa: the 16 literal labels are inherited from the earlier Swedish direct-pollen cohort.
The exact axis is a closed subcomposition. A missing literal Neotoma NISP row is treated as an unreported zero cell under the documented Neotoma count-matrix workflow. All remaining pollen labels are outside the response and are not interpreted as absent vegetation. Coordinates for source polygon geometries use their WGS84 geometric centroids; source point geometries are retained as points.
| Held-out country | Candidate samples | Candidate sites | Candidate bins | Positive 16-label bins | Predicted bins | Scored bins | Refused bins |
|---|---|---|---|---|---|---|---|
| DK | 91 | 1 | 24 | 24 | 24 | 24 | 0 |
| FI | 1797 | 28 | 435 | 435 | 435 | 435 | 0 |
| NO | 3544 | 52 | 1178 | 1177 | 1084 | 1084 | 94 |
| Country | Method | Attempted predictions | Scored bins | Scored sites | Site-macro MAE | Site-macro Hellinger |
|---|---|---|---|---|---|---|
| DK | training_mean | 24 | 24 | 1 | 0.037484 | 0.423937 |
| DK | nearest_site | 24 | 24 | 1 | 0.029809 | 0.353519 |
| DK | inverse_distance_squared | 24 | 24 | 1 | 0.033022 | 0.387704 |
| DK | clr_rbf_ridge | 24 | 24 | 1 | 0.039220 | 0.393744 |
| FI | training_mean | 435 | 435 | 28 | 0.038799 | 0.338505 |
| FI | nearest_site | 435 | 435 | 28 | 0.033223 | 0.289538 |
| FI | inverse_distance_squared | 435 | 435 | 28 | 0.034375 | 0.306323 |
| FI | clr_rbf_ridge | 435 | 435 | 28 | 0.082477 | 0.620855 |
| NO | training_mean | 1084 | 1084 | 52 | 0.030190 | 0.287670 |
| NO | nearest_site | 1084 | 1084 | 52 | 0.032280 | 0.327944 |
| NO | inverse_distance_squared | 1084 | 1084 | 52 | 0.027012 | 0.260268 |
| NO | clr_rbf_ridge | 1084 | 1084 | 52 | 0.034496 | 0.323876 |
Site-macro scores average bins within each physical site and then weight sites equally within country. Denmark has one physical site group and cannot estimate between-site Danish variation. Country-held-out scores are descriptive transfer results from this captured EPD subset, not evidence for all Nordic landscapes or all time periods. coverage.csv retains site-level age/count candidate bins and separates positive closed-axis outcomes, predicted bins, scored bins, empty/refused target outcomes, and eligible bins without a prediction.
Source and limits¶
The Danish, Finnish, Norwegian and Swedish cohorts come from European Pollen Database datasets distributed through Neotoma. The Neotoma source review records EPD reuse under CC BY 4.0 with Neotoma, EPD, and original investigator attribution. Dataset identifiers, DOIs, and contributors are retained in cohort.csv and training-sources.csv. No PPE, fall-speed, basin-area, or vegetation-reference assumption is needed for observed-pollen prediction; those inputs remain blocked for LRA and measured-vegetation claims.
All selected candidates have one source-owned comparable sample-age claim after the declared chronology rule: when several comparable claims exist, use the unique source-flagged default for its age type; otherwise refuse the sample. Neotoma permits one default chronology per age type, so a radiocarbon and a calibrated chronology can both be flagged without contradicting each other (Neotoma chronology manual). The corrected source projection removes false cross-age-type conflicts while retaining genuine same-age-type conflicts. Sample age type labels remain in the input profiles. Source uncertainty bounds vary by record and are not propagated into the exploratory fixed 500-year bins.
Target bins lacking two positive-count training sites, bins with no positive count on the fixed 16-label axis, and predictions without target outcomes remain unscored; the CSV files retain the actual prediction and scoring denominators. Spatial coordinates from polygon geometry centroids are approximate source locations. No independent vegetation truth was used.
Reproduction¶
From the repository root, run PYTHONPATH=packages/bijux-pollenomics/src:packages/bijux-pollenomics-dev/src .venv/bin/python -m bijux_pollenomics_dev.source_completion.nordic_direct_pollen prepare --relational-root data/neotoma/relational --raw-root data/neotoma/raw/neotoma_pollen_dataset_downloads --workdir artifacts/nordic-direct-pollen, then run predict separately for DK, FI and NO, then run score with --relational-root data/neotoma/relational --workdir artifacts/nordic-direct-pollen --publish-dir docs/report/regions/nordic/direct-pollen. The prediction commands do not read the held-out outcome inputs. Input and prediction hashes are retained in manifest.json.