Wavelength-Aware Local-Frame Attention Retrieval Network for Hyperspectral Mineral Identification
Abstract
Hyperspectral mineral identification is essential for geological remote sensing and alteration mapping. However, existing spectral matching, closed-set classification, and metric-learning methods often fail to fully exploit wavelength-dependent local absorption morphology and intraclass spectral variability. This article proposes the wavelength-aware local-frame attention retrieval network (WAFER-Net), a retrieval-verification model for mineral spectral recognition. WAFER-Net divides full reflectance spectra into overlapping local spectral frames and jointly encodes local absorption shape, derivative variation, spectral-shape statistics, and frame-center wavelength position. A smooth sparse frame-attention module emphasizes diagnostic absorption regions, while a global-trend branch preserves broad-band continuum information. Class-conditional nearest-neighbor retrieval in an L2-normalized embedding space further preserves support-sample-level variability. On a 54-class United States Geological Survey (USGS) mineral spectral-library dataset, WAFER-Net achieves mean Top-1 overall accuracy (OA), Top-3 accuracy, Top-5 accuracy, and mean reciprocal rank values of 0.9641 ± 0.0064, 0.9871 ± 0.0016, 0.9935 ± 0.0013, and 0.9767 ± 0.0041, respectively, over five random seeds. In the airborne visible/infrared imaging spectrometer-cuprite experiment, resampled USGS spectra and 174 independently identified AVIRIS spectra are jointly modeled, yielding an OA of 0.9378, Macro-F1 of 0.9332, balanced accuracy of 0.9366, and scene-level agreement of 0.9122 with the public reference map. These results show that joint modeling of local spectral shape, wavelength position, and support-sample matching improves fine-grained mineral recognition and real-scene mapping stability.