Reproducibility data and code for Spectral Patterns and Reduced-Band Neural Learning in UAV Hyperspectral Images of a Landfill

Reproducibility materials for “Spectral Patterns and Reduced-Band Neural Learning in UAV Hyperspectral Images of a Landfill”.The study uses hyperspectral images from the 14 April 2026 campaign at Los Barrios, Spain. It groups surface spectra without using external references in the grouping and evaluates whether neural networks can approximate a spectral-reference rule using fewer input bands than the rule itself. Earlier TDLAS measurements provide methane context, not pixel-level labels. Results measure spectral-rule learning, not validated methane detection.The package includes extracted samples, group centres, digitized reference curves with provenance, native-resolution masks, campaign records, image headers, fixed training/validation/test designs, numerical code, trained weights, predictions and verification records. It contains 112 main and secondary neural runs and 24 additional SWIR neural runs using a 3 m geographic exclusion on the same test zones, plus logistic and direct controls. Mean MLP AUC is 0.917 at 1 m and 0.915 at 3 m.Additions for the second review round include a sensitivity analysis of the labelling rule comprising the nominal rule and 21 perturbations, the corresponding labels and results, a 272-band sample cache, complete paired results by zone, and a documented chronology of the protocol records. The original data, nominal labels, training designs, trained weights and predictions are preserved. Networks were not retrained for the sensitivity analysis; archived predictions were evaluated against the alternative labels. Version 1.18 updates documentation and verification records without changing the scientific files or results of version 1.17.Technical Documentation provides implementation settings, complete counts, results by zone, sensitivity results and scene maps. Compact inputs support the documented recalculations and neural retraining. Full-scene extraction and clustering require ten complete processed reflectance products (69,748,089,728 bytes), inventoried but not included; original detector recordings are not supplied. The README explains these reproducibility limits and provides verification commands. External digitized curves and mineral data retain their source terms.