Best candidate per species, or an explicit refusal

A species with no scorable reference series MUST NOT produce a recommendation. Three of the six species here are in that position: amabilis fir, subalpine fir and hybrid spruce have too few ground plots to bin, and their SORTIE curves are marked available_unused in the project's curve-selection table because that model is not well parameterized for them in the ICH. Ranking hundreds of indistinguishable all-NA rows would return whichever combination happened to sort first, dressed up as a result.

Usage

growth_best_candidates(
  scores,
  growth_params,
  windows,
  scoring = NULL,
  top_frac = 0.1,
  level_frac_warn = 0.9
)

Arguments

scores

A tibble from growth_score_fit() with an objective_rmse.

growth_params

The parameters currently in use.

windows

A tibble from growth_fitting_windows().

scoring

A tibble from read_growth_scoring(), or NULL.

top_frac

Numeric. Fraction of each species' ranking treated as indistinguishable from the winner, over which the level band is reported.

level_frac_warn

Numeric. Warn when the selected candidate reached less than this fraction of its own asymptote, so its level is an extrapolation. NA disables the warning.

Value

One row per species.

Details

Those species come back with fitted = FALSE, no parameters, and the values currently in use carried through untouched. That is the honest answer, and it names what would change it: more plots, or promoting a SORTIE curve to used.

Reading the level band

Ranking is on shape alone and the level is recovered afterwards, by dividing the reference plateau by the fraction of its own asymptote the simulated curve reached (see growth_inflation_factor()). That division is the whole reason biomass_max_est is not simply read off the winner and trusted: its leverage on any error in the simulated curve is 1 / achieved_frac. A candidate that plateaued cleanly recovers its level almost exactly, while one still climbing when the run ended is extrapolating, and two candidates a fraction of a percent apart in shape error can then imply levels differing by a factor of two or more.

So biomass_max_lo / biomass_max_hi report the range of biomass_max_est across the candidates that cannot be told apart from the winner, and level_extrapolated flags a winner that never approached its asymptote. A wide band is not noise to be averaged away: it says the references do not determine the level, and the fix is a longer run, a better-constrained reference, or a wider sweep – not a different summary of the same surface. Use growth_identifiability() to see which swept parameter is responsible.

See also

Other growth calibration helpers: extract_landscape_cohort_structures(), growth_add_objective(), growth_auto_window(), growth_bin_observations(), growth_calibration_design(), growth_calibration_partition(), growth_calibration_work_root(), growth_climatic_distance(), growth_climatic_weight(), growth_expand_over_pseudo_species(), growth_factorial_ratio_grid(), growth_fitting_windows(), growth_identifiability(), growth_inflation_factor(), growth_mortality_onset_frac(), growth_plot_palette(), growth_pseudo_species_name(), growth_reference_curves(), growth_reference_inflection(), growth_score_fit(), growth_scoring_for(), growth_smooth_observations(), growth_structure_cell_curves(), growth_structure_cohort_table(), growth_structure_design(), growth_structure_summary(), growth_window_for(), plot_growth_calibration(), plot_growth_candidate(), plot_growth_factorial_sensitivity(), plot_growth_structures(), read_growth_scoring(), read_landscape_cohort_structures(), scale_linetype_growth_reference(), write_growth_review_bundle()