Optimal Hyperspectral Undersampling Strategy for Satellite Imaging

Vita V. Vlasova, Vladimir G. Kuzmin, Maria S. Varetsa, Natalia A. Ibragimova, Oleg Y. Rogov, Elena V. LyapuntsovaarXiv 2025 · OpenReview

TL;DR

This paper introduces Iterative Wavelet-based Gradient Sampling (IWGS), a novel method for selecting the most informative spectral bands in hyperspectral images. IWGS significantly improves classification accuracy and computational efficiency, making it ideal for resource-constrained applications like edge devices.

What Problem Does This Paper Address?

Hyperspectral imaging (HSI) captures incredibly detailed spectral information, enabling precise analysis of Earth's surface. However, this wealth of data comes with a major challenge: high dimensionality. This means a massive number of spectral bands, leading to increased computational costs, redundancy, and difficulties in training accurate classification models, especially when labeled data is scarce. Traditional methods often struggle to balance selecting informative bands with preserving crucial spectral details, particularly for minority classes. This paper aims to tackle this by developing a more intelligent and efficient band selection strategy.

Methods

The core of this work is the Iterative Wavelet-based Gradient Sampling (IWGS) algorithm. Unlike traditional band selection methods, IWGS operates in the wavelet domain. Here's how it works:

  • Wavelet Transformation: The hyperspectral data is first transformed into the wavelet domain. This process decomposes the signal into different frequency components, effectively highlighting spectral features and introducing sparsity.
  • Iterative Gradient Sampling: IWGS iteratively selects bands. In each step, it evaluates how much selecting a particular band would reduce the classification loss of a model. This is done by computing gradients in the wavelet domain.
  • Task-Specific Selection: The selection is "task-specific," meaning it directly optimizes for better classification performance on the given data and model.
  • Robustness: By operating in the wavelet domain and using an iterative approach, IWGS aims to be more robust to noise and preserve subtle spectral variations crucial for distinguishing between similar classes.
  • Hybrid Architecture: The paper also proposes a hybrid neural architecture that combines CNNs for spatial feature extraction with a lightweight attention-inspired spectral modeling approach, drawing inspiration from State Space Models (SSMs) like Mamba for efficient long-range dependency modeling.

Datasets

The proposed methods were evaluated on two widely-used benchmark hyperspectral image datasets:

  • Houston 2013: Used for evaluating the IWGS algorithm's accuracy and efficiency.
  • Indian Pines: A well-known dataset acquired by the AVIRIS sensor, consisting of 145x145 pixels with 220 spectral bands. This dataset is particularly challenging due to high spectral similarity between classes, making it ideal for testing the model's ability to handle class imbalance and nuanced separation.

Key Insights and Ideas

  • Wavelets for Smarter Band Selection: Transforming data into the wavelet domain is a powerful way to reveal spectral features and reduce redundancy. The gradients in this domain provide a more insightful signal for selecting informative bands compared to raw spectral data.
  • Iterative Optimization is Key: The iterative nature of IWGS allows it to adaptively refine band selection, progressively discarding less useful bands and focusing on those that truly improve classification.
  • Balancing Accuracy and Efficiency: The proposed IWGS strategy offers a significant improvement in both classification accuracy and computational efficiency, making it a practical solution for real-world applications where resources are limited.
  • Patch Size Matters: The study highlights that the choice of spatial patch size for analysis is crucial and has a significant impact on classification performance, with different classes benefiting from different patch sizes.

Key Results

  • High Accuracy: IWGS achieved an overall accuracy of up to 97.8% on the Indian Pines dataset for selected classes. On the Houston 2013 dataset, it achieved 97.60% accuracy.
  • Computational Efficiency: The method significantly reduces computational demands and memory footprint compared to conventional architectures, making it suitable for edge devices.
  • Robustness: IWGS demonstrated robustness to atmospheric noise and adversarial perturbations, maintaining high performance even under challenging conditions.
  • Class Balance: The sampling process effectively preserved spectral diversity and class-specific structure, leading to more balanced training data and improved performance on underrepresented classes.
  • Optimal Trade-off: For practical applications, a patch size of 5 was identified as an optimal balance between accuracy and computational efficiency, offering 97.10% OA.

Limitations and Future Work

  • Parameter Sensitivity: The paper mentions conducting a parameter sensitivity analysis, focusing on spatial patch size. While optimal configurations were found, specific applications might benefit from further fine-tuning.
  • Future Work: The authors suggest that the convergence of CNNs, transformers, and SSMs is expected to yield increasingly powerful models. Their work contributes to this by offering a more accessible hyperspectral analytics solution. Further research could explore integrating IWGS with more advanced deep learning architectures or applying it to other types of spectral data.

Summary updated

Abstract

Hyperspectral image (HSI) classification presents significant challenges due to the high dimensionality, spectral redundancy, and limited labeled data typically available in real-world applications. To address these issues and optimize classification performance, we propose a novel band selection strategy known as Iterative Wavelet-based Gradient Sampling (IWGS). This method incrementally selects the most informative spectral bands by analyzing gradients within the wavelet-transformed domain, enabling efficient and targeted dimensionality reduction. Unlike traditional selection methods, IWGS leverages the multi-resolution properties of wavelets to better capture subtle spectral variations relevant for classification. The iterative nature of the approach ensures that redundant or noisy bands are systematically excluded while maximizing the retention of discriminative features. We conduct comprehensive experiments on two widely-used benchmark HSI datasets: Houston 2013 and Indian Pines. Results demonstrate that IWGS consistently outperforms state-of-the-art band selection and classification techniques in terms of both accuracy and computational efficiency. These improvements make our method especially suitable for deployment in edge devices or other resource-constrained environments, where memory and processing power are limited. In particular, IWGS achieved an overall accuracy up to 97.8% on Indian Pines for selected classes, confirming its effectiveness and generalizability across different HSI scenarios.
BibTeX
@misc{vlasova2025optimal,
  author = {Vita V. Vlasova and Vladimir G. Kuzmin and Maria S. Varetsa and Natalia A. Ibragimova and Oleg Y. Rogov and Elena V. Lyapuntsova},
  title = {{Optimal Hyperspectral Undersampling Strategy for Satellite Imaging}},
  year = {2025},
  eprint = {2504.19279},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.19279},
}
Figure