Spectral-Space-Aware Annotation and PCA-Assisted Quality Control for Hyperspectral Datasets

Reliable annotation and quality control remain key bottlenecks in hyperspectral data analysis because of high spectral dimensionality, intra-class spectral variability, and the limited spectral-validation capacity of conventional labeling tools. Most annotation environments were originally developed for RGB imagery and therefore lack wavelength-aware visualization, spectral inspection, and spectral-space-aware label validation, which are essential for hyperspectral workflows.

This study introduces an interactive hyperspectral annotation and PCA-assisted quality-control framework that integrates cube visualization, multi-mode ROI annotation, wavelength-aware spectral inspection, ROI-linked spectral comparison, and dimensionality-reduction-based label validation within a single environment. The framework supports Polygon, Point, Rectangle, Brush, and Eraser annotation modes. Its central contribution is the direct integration of PCA-based spectral-space validation into the annotation workflow. Each ROI is represented by a spectral vector, projected into a low-dimensional PCA space, and screened for spectrally inconsistent behavior using a robust Median Absolute Deviation criterion. Flagged outliers remain interactively linked to their original image-space locations, enabling immediate inspection and removal.

Quantitative evaluation confirmed measurable improvements in annotation consistency. From an initial set of 29 ROIs, 4 spectrally inconsistent annotations were detected and removed, resulting in an 86.2% retention rate. The mean PCA-space distance to the class centroid decreased from 19.95 to 10.54, corresponding to a 47.2% reduction in cluster dispersion. Mean spectral variance across wavelengths decreased from 0.0107 to 0.00247, representing a 76.8% reduction in spectral variability. The first two principal components explained approximately 79.7% of the total spectral variance, and evaluation across six hyperspectral samples produced retention rates of 79.4–94.6%. These results demonstrate that integrating spectral-space-aware QC directly into the annotation stage improves dataset compactness, annotation reliability, and generation of machine-learning-ready hyperspectral datasets. Future work will include comparative and ablation studies using alternative outlier-detection criteria, dimensionality-reduction methods, and larger multi-class datasets.