Subset-Driven Representation Learning for Hyperspectral Image Classification: Assessing Latent Space Quality Through Spectral–Spatial and Transformer-Based Models

Quirita V.A.A.; Quirita V.H.A.; Aliaga W.; Costa G.A.O.P.; Cardenas J.D. · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2026) · DOI: 10.1109/jstars.2026.3727184

Resumen

Hyperspectral representation learning commonly relies on complete scenes, increasing computational cost even when only compact latent features are required for classification. This work evaluates whether reduced-scene pretraining can preserve discriminative latent spaces for hyperspectral image classification. We compare PCA, AE, AE-Patch, VIT, MAE-VIT, and DINO-VIT across multiple training fractions (k), latent dimensions (d), downstream classifiers, and five benchmark datasets: Pavia University, Indian Pines, Salinas Valley, Kennedy Space Center, and WHU-Hi-LongKou. Results show that comparable latent representations can often be learned using reduced training fractions, yielding competitive macroaveraged F1 scores, especially when spectral–spatial or self-supervised transformer representations are used. The analysis further suggests that, within the evaluated configurations, changes in latent dimensionality are often associated with clearer performance variations than changes in the pretraining fraction, while full-scene pretraining produces only modest classification performance gains despite substantially higher latency, VRAM usage, and FLOPs. Dataset structure also modulates this behavior, with Pavia University and WHU-Hi-LongKou showing greater robustness to representation-learning data reduction, whereas Indian Pines, Salinas, and Kennedy Space Center exhibit larger class-wise variations due to spectral overlap and class imbalance. Overall, subset-driven representation learning offers a practical F1 score–cost tradeoff for scalable hyperspectral classification.

Instituciones: Pontificia Universidad Católica del Perú; Universidad del Pacífico