Addressing Limitations of Common Methods in Attention-Based Hyperspectral Band Selection Algorithms

Hyperspectral band selection algorithms are recognized for their ability to reduce dimensionality, improve accuracy, and enhance computational efficiency. Recently, several deep learning-based algorithms with attention mechanisms have been proposed to weight and rank bands for selection. The efficacy of these algorithms depends on internal methods that balance comprehensive assessment of all bands with downweighting unimportant ones. These methods include weight aggregation across data samples and batches, weight normalization, and recalibrated data processing through a suitable downstream task. In this study, we interrogated the fundamental suitability of these methods and demonstrated that those suited for typical deep learning tasks may not be the best fit, as the primary objective here is specific to band ranking rather than accurate inference on the downstream task. Based on the issues identified, we propose an algorithm that addresses these limitations by properly aggregating weights derived across samples and data batches, normalizing weights without hindering gradient flow, processing recalibrated data based on the altered characteristics, and ensuring proper assessment before discarding any band. Our evaluation on three baseline datasets shows that our algorithm consistently outperforms both traditional and recent algorithms, achieving up to a 5% improvement in overall accuracy and a 7% improvement in average accuracy.