Beyond Accuracy: A Multi-Axis Evaluation Framework for Interpretable Topic Models on Hyperspectral Imagery

Flagship paper (programme paper P1) of the CAOS_LDA_HSI series. Latent Dirichlet Allocation (LDA) has been adapted to hyperspectral imagery (HSI) as a route to interpretable spectral mixtures: each spectral unit (canonically a pixel) becomes a document of quantised band-frequency tokens and the inferred topic-word distributions act as non-negative basis spectra. Yet the dominant evaluation criterion remains downstream classification accuracy, a view that is silent on whether the basis is reproducible, band-robust, calibrated, or actually outperforms standard linear baselines. We propose and instantiate a twelve-axis evaluation framework (F-1 hierarchical-Bayesian accuracy over five method families; F-2 topic-word coherence; F-3 seed stability for LDA, ProdLDA and ETM; F-4 capacity sensitivity; F-5 band-mask robustness; F-6 cross-method clustering agreement; F-7 topic-label coupling; F-8 Hungarian-aligned identity tracking; F-9 HIDSAG cross-preprocessing stability; F-10 cross-scene transfer; F-11 rate-distortion; F-12 external validation). We apply it to six standard HSI benchmarks (Indian Pines, Salinas, Salinas-A, Pavia University, Kennedy Space Center, Botswana) and five HIDSAG mineralogical subsets. Three findings the accuracy-only view does not surface: (1) the topic-routed soft method (0.741 +/- 0.193) is statistically indistinguishable from the raw-spectrum logistic baseline (0.736 +/- 0.193, P[A>B]=0.64), so the topic representation buys interpretability without losing accuracy; (2) band-mask ARI separates band-robust scenes (Salinas-A SWIR-only, paired ARI 0.766) from non-robust ones (KSC, Botswana, ARI ~0.01); (3) on HIDSAG the topic-logistic posterior (0.316 +/- 0.220) falls well below raw-logistic (0.585 +/- 0.221), inverting the labelled-scene picture. We release 1734 deterministic derived artefacts (449 MB), the builder source, the FastAPI backend, the React frontend and a 109-endpoint smoke harness as a public, MIT-licensed reproducibility package. Code and derived artefacts: https://github.com/fsantibanezleal/CAOS_LDA_HSI . Interactive web application: https://lda-hsi.fasl-work.com . Manuscript sources: https://github.com/fsantibanezleal/CAOS_LDA_HSI_Paper . Funding: The Advanced Mining Technology Center (AMTC) Basal project (ANID/PIA Project AFB220002) and ANID FONDECYT Postdoctorado 3220094.