Plant Spectral Quantiles¶
Module: spectralbridge.plant_spectral_quantiles.
Checkpointed boosted conditional quantile signatures and held-plant matching.
Fitted curves are representations, never synthetic independent plant observations. Species fits use independent polygon medians; individual fits use repeated-flight polygon medians. Unsupported NEON segments are trained separately.
API glossary · Workflow / scientific context
Signatures and documentation below are collected from the current source. Advanced objects are not newly declared stable by inclusion here.
Audience: advanced.
QuantileModelConfig
dataclass
¶
QuantileModelConfig(quantiles: tuple[float, ...] = QUANTILES, boosting_rounds: int = 80, learning_rate: float = 0.1, max_depth: int = 5, max_bin: int = 256, seed: int = 0, min_plants_per_species: int = 10, holdout_fraction: float = 0.2, workers: int = 4, stages: tuple[str, ...] = _STAGES)
Configure historical boosted quantile spectral models, group eligibility and rendering budgets.
Audience: advanced.
fit_quantile_signature ¶
fit_quantile_signature(wavelengths, spectra, *, config=None, hyperspectral=True)
Fit exact supplied wavelengths with equal total weight per input spectrum.
NaN columns and >8 nm hyperspectral gaps split independent boosting models. Returns raw fitted quantiles and explicitly pointwise-rearranged ordered quantiles. No clipping, normalization, extrapolation or added samples.
Audience: advanced.
quantile_distance ¶
quantile_distance(observed, curves, quantiles, *, minimum_bands=2)
Mean pinball loss times two: a discrete 1–99% integrated-score approximation.
Lower is a closer match in supplied units. It is not a class probability and treats wavelengths marginally, without estimating spectral covariance.
Audience: advanced.
rank_unknown_spectrum ¶
rank_unknown_spectrum(output_dir, wavelengths_nm, values, *, stage='corr', region=None, level='species', exclude_plant_ids=(), minimum_shared_fraction=0.5, allow_unverified_export_coordinates=False)
Rank supplied native-unit spectra against full quantile references.
Pointwise pinball ranking is exploratory; scores are not identification probabilities. Query gaps/exclusions are preserved, with no extrapolation. Individual references represent repeated observations of one polygon.
Audience: advanced.
run_boosted_quantile_atlas ¶
run_boosted_quantile_atlas(comparisons_dir, output_dir, *, config=None, make_individual_pdf=True)
Fit species/site and individual-polygon quantile signatures with checkpoints.
All 1–99% quantiles are jointly trained with separate quantile tree outputs. Uses completed compact tables only. Models/plots are derivatives; original pixel library and parent plant tables remain immutable.