Scholarlink

The registration of hyperspectral (HS) images, defined as the alignment of the same pixels in different HS images, is an essential preliminary step to enhance the spectral information available for each pixel in HS image analysis. Such alignment enhances various HS applications including change detection, temperature emissivity separation, noise suppression, and target classification across different spectral ranges. After discussing the need for HS image registration, this chapter presents the main class of methods in the literature with respect to the spectral region of the HS image pairs and the utilized technical methodology. The chapter first reveals that the majority of registration studies are concentrated on HS images in the reflectance band that encompasses VNIR (0.4–1 µm) and SWIR (1–2.5 µm) regions. The main technical methods in these regions are categorized as key point based, optimization based, and deep learning based methods. Key point-based methods extract and match key-points on HS image pairs to estimate the geometric transformation between them, whereas optimization-based methods maximize a closeness metric between the transformed and reference images by iteratively updating a randomly initialized geometric transformation between the pairs. The third category, deep learning-based methods, performs the same operation by training different types of networks, including autoencoder and spatial transformer networks, to learn the registration directly from the data. In thermal MWIR (3–5 µm) and LWIR (7.5–14 µm) regions, newly developing registration methods show satisfactory performance for image pairs captured on the same day. However, the performances decrease for images captured on different days due to thermal changes. The future challenges for HS image registration include developing new key point descriptors for cross-registration of reflectance and thermal band HS images, addressing the lack of labeled data for deep learning-based methods, and the absence of experimental analysis to evaluate the impact of registration on target classification performances. Addressing these challenges will be crucial for advancing HS image registration and its applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hyperspectral Image Registration: State-of-the-Art and Future Challenges

  • 章节

  • 发表日期：

    2025年

查看全文
  • Alper Koz

摘要

The registration of hyperspectral (HS) images, defined as the alignment of the same pixels in different HS images, is an essential preliminary step to enhance the spectral information available for each pixel in HS image analysis. Such alignment enhances various HS applications including change detection, temperature emissivity separation, noise suppression, and target classification across different spectral ranges. After discussing the need for HS image registration, this chapter presents the main class of methods in the literature with respect to the spectral region of the HS image pairs and the utilized technical methodology. The chapter first reveals that the majority of registration studies are concentrated on HS images in the reflectance band that encompasses VNIR (0.4–1 µm) and SWIR (1–2.5 µm) regions. The main technical methods in these regions are categorized as key point based, optimization based, and deep learning based methods. Key point-based methods extract and match key-points on HS image pairs to estimate the geometric transformation between them, whereas optimization-based methods maximize a closeness metric between the transformed and reference images by iteratively updating a randomly initialized geometric transformation between the pairs. The third category, deep learning-based methods, performs the same operation by training different types of networks, including autoencoder and spatial transformer networks, to learn the registration directly from the data. In thermal MWIR (3–5 µm) and LWIR (7.5–14 µm) regions, newly developing registration methods show satisfactory performance for image pairs captured on the same day. However, the performances decrease for images captured on different days due to thermal changes. The future challenges for HS image registration include developing new key point descriptors for cross-registration of reflectance and thermal band HS images, addressing the lack of labeled data for deep learning-based methods, and the absence of experimental analysis to evaluate the impact of registration on target classification performances. Addressing these challenges will be crucial for advancing HS image registration and its applications.