SALHIC: An Active Learning-Inspired Framework for Supervised Hyperspectral Image Classification

Quirita V.H.A.; Quirita V.A.A.; Costa G.A.O.P.; Miguel A.P. · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2026) · DOI: 10.1109/jstars.2026.3728312

Resumen

Hyperspectral image classification (HIC) remains challenging due to high spectral dimensionality, spectral ambiguity, class imbalance, and the limited availability of labeled data. In this work, we propose SALHIC, a supervised iterative refinement framework inspired by active learning principles for HIC. Unlike conventional active learning approaches, SALHIC does not require unlabeled data or oracle-based annotation. Instead, it iteratively updates a neural network-based classifier by incorporating informative labeled observations selected from the training set according to its prediction uncertainty. The framework was evaluated on four hyperspectral datasets (Indian Pines, Pavia University, Kennedy Space Center, and H18) under both pixel-wise and patch-wise classification settings and under two experimental protocols representing moderate to severe label scarcity. Across the resulting 16 experimental scenarios, SALHIC improved upon the strongest baseline by up to 32.43 percentage points in macroaveraged F1-score, with the largest gains consistently obtained under severe label scarcity. A class-wise analysis further showed that the improvements concentrate on categories that remain poorly discriminated after conventional supervised training rather than on minority classes. The results indicate that SALHIC is most effective when the baseline leaves measurable room for improvement and the scene contains spectrally similar categories whose ambiguity remains unresolved after conventional training.

Instituciones: Pontificia Universidad Católica del Perú; Universidad del Pacífico