A Band-Mask Robustness Diagnostic for Latent Dirichlet Allocation on Hyperspectral Imagery

Latent Dirichlet Allocation (LDA) and its neural variants are increasingly used as interpretable spectral mixture models on hyperspectral imagery. Their seed and capacity stability has been studied at length; their robustness to the choice of spectral window has not. We introduce a band-mask robustness diagnostic that refits a canonical LDA model under four band-restriction policies (VNIR-only, SWIR-only, atmospheric-water-band removal, top-50 Fisher discriminant) and compares the masked dominant-topic maps against the canonical fit via the adjusted Rand index after a Hungarian topic-id alignment. Applied to six standard hyperspectral scenes (Indian Pines, Salinas, Salinas-A, Pavia U, Kennedy Space Center, Botswana) and five HIDSAG mineral subsets, the diagnostic separates the scenes: topic identities on Salinas-A under SWIR-only restriction recover the canonical assignment with paired ARI = 0.77, while KSC paired ARI is approximately 0.01 under every mask, and Botswana falls to approximately 0.01 only under SWIR (elsewhere 0.11-0.13). The diagnostic therefore separates scenes on which any interpretable-spectral-mixture claim is band-robust from scenes on which it is not. We release the full sweep (44 LDA refit attempts, 43 successful) as a public, deterministic artefact set under a permissive licence to support follow-on work. 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.