How do we know SpectralBridge works?¶
Evidence has several levels. Correct mathematics and complete files are necessary; neither proves a broad ecological application works.
| Level | Evidence | What remains outside that claim |
|---|---|---|
| 1. Mathematical correctness | Analytic spectra/fractions, independent constrained optimizer, QA/scaling/geometry tests | Real-world measurement accuracy |
| 2. Pipeline contracts | IO, orientation, ancillary alignment, chunking, filename contracts, restart and source immutability | External sensor agreement |
| 3. Biological association | Whole-curve conditional permutation analysis; recorded researcher cover categories only | Causal functional mechanisms, novel-species classification |
| 4. Synthetic sensor reduction | Paired memberships/history, effect/direction/rank retention in OLI | Independent satellite or drone observations |
| 5. Real-sensor comparison | One independently selected next-day R10C/OLI pilot with QA, footprint and uncertainty checks | Multisite/general sensor interchangeability |
| 6. Mixture/abundance inference | Experimental synthetic fraction recovery and ambiguity tests | Operational ecological fraction accuracy; incomplete independent truth |
Read execution versus QA evidence, whole-spectrum methods, real Landsat and MESMA readiness for the detailed scopes. The test hierarchy explains what automated tests protect. Canonical reports keep full memberships, seeds, hashes and failures; website figures summarize rather than replace them.
Known limitations¶
Library physical scaling remains unresolved; use stored units until source metadata confirms a conversion. Multiple wavelength grids and structural NaNs must not be zero-filled into a universal spectrum. Species/PFT and environmental composition are confounded. Repeated plants, years, overlapping contrasts and shared source spectra are dependent. MicaSense physical coordinates remain blocked in the current library export. Synthetic OLI shares its parent NEON source. The real pilot has a next-day lag and legacy timezone caveat. Independent pure-material and full fractional-cover truth are missing.
Passing a test or threshold does not eliminate these limitations. New geographic, sensor or ecological applications require validation designed for that target.
Integrated information-retention audit¶
The 2026-10-14 integrated audit connects biological differences to whole-plant, acquisition, year and site holdouts, training-only wavelength selection, prototype coverage, withheld-class rejection and calibration. It reuses the canonical functional, sensor and mixture evidence rather than treating classification as abundance validation. Full-spectrum logistic recognition is strong for held-out plants in this sample; site transfer is uneven and unknown-class rejection weak.
The experimental workflow requires the optional analytics extra. Its staged
entry point is examples/spectral_library_analytics.py; --stage profile measures
proper-training cost, while explicit learning/extensions/domains/mixtures stages
run the declared experiments. --stage all completes secondary diagnostics before
sealing the learning ledger. Training-only preprocessing, GlobalID purging and
independent calibration are required. This is not a stable top-level package API.
The source-specific human PDF, editable narrative, master tables, code-linked
model index, figures and checks are recorded under
provenance/spectral-library-analytics/2026-10-14/. The report is generated at
output/pdf/spectral_library_analytics/SpectralBridge_Spectral_Library_Analytics_Report.pdf.
Read its evidence dimensions and limitations before drawing a deployment claim.