Preface

Hyperspectral imaging (HSI) is an aerial imaging technology that measures the way an object reflects and emits light at different wavelengths. Typically, it can cover hundreds of bands of light in the electromagnetic spectrum, revealing the precise spectral properties of materials found in the region of interest. With the resulting data, the methodology can distinguish the subtle differences between similar objects, allowing it to map out and differentiate objects and materials in great detail. Due to its fine-grained resolution and ability to distinguish different chemical species, HSI is becoming a powerful tool to spatially resolve the chemistry of materials in varying scientific and engineering disciplines.

HSI data acquisition involves the use of an aerial detector multiplexed in two dimensions and, therefore, requires multiple measurements to complete one data acquisition cycle. The multiple measurements can be executed in two different ways, position scanning or wavelength scanning. Position scanning HSI acquires 2D data of one spatial dimension and the spectral dimension, and scans across the other spatial dimension, whereas wavelength scanning HSI multiplexes the two spatial dimensions and scans across the spectral dimension. Clearly, both methods need time to complete a data cube acquisition. Enabling fast HSI will open doors to new applications where multiple constituents or spatiotemporal dynamics need to be resolved. A variety of snapshot techniques have been developed by invoking a spatial-spectral modulation scheme, such as illuminating an object with a coded light pattern or inserting a spectral modulation module in the HSI imaging device.

HSI is currently applied in many fields. However, we also face a new challenge in data processing and in how to reliably retrieve meaningful information from the highdimensional HSI data cubes in real-time. Recent progress in both machine-learning and deep-learning techniques may offer a solution to this issue. This book brings together a collection of five chapters offering a glimpse of the status of machine- and deep-learning methodological development for hyperspectral imaging applications.

Chapter 1 "Perspective Chapter: Hyperspectral Imaging for the Analysis of Seafood", by Samuel Ortega et al., presents a survey of current uses of hyperspectral technology for seafood evaluation. The authors briefly describe the optical properties of tissue and offer an introduction to the instrumentation and the developmental status of HSI in the relevant aspects of the seafood industry.

As noted above, consistent data preprocessing and reliable feature extraction are the first step to meaningful information retrieval from high-dimensional data cubes. Chapter 2, "Useful Feature Extraction and Machine Learning Techniques for Identifying Unique Pattern Signatures Present in Hyperspectral Image Data", by Jeanette Hariharan et al., presents a data preprocessing protocol for HSI data. The authors review feature extraction techniques that are useful for identifying pattern

signatures embedded in hyperspectral data, and discuss the best practices for processing and analyzing hyperspectral data using machine-learning techniques.

Accurate recording of HSI data ideally requires the data to be acquired with high spatial and spectral resolution. Although until now this has been a fairly lengthy process, deep-learning techniques have recently been developed to improve it. Chapter 3, "Unsupervised Deep Hyperspectral Image Super-Resolution", by Zhe Liu and Xian-Hua Han reviews recent advances in deep unsupervised frameworks for generating high-resolution (HR) HSI, demonstrating with a universally learnable module that only uses low-quality observations to reconstruct the underlying HR-HSI. K. Priya and K.K. Rajkumar, the authors of Chapter 4, "Hyperspectral and Multispectral Image Fusion Using Deep Convolutional Neural Network - ResNet Fusion", point out that in a convolutional neural network (CNN), each layer takes the output from the previous layer, and tends to lose information as the network goes deeper into the architecture. They implement a fusion process in a Residual Network (ResNet) by adding the skip connection between the convolution layers. This skip connection helps to extract more detailed features from the images without any information degradation. The authors measured the results of their ResNet fusion method and found that it exhibits outstanding performance compared with all traditional methods.

In Chapter 5, "Magnetic Scattering with Polarised Soft X-rays", Paul Steadman and Raymond Fan offer an indication of the direction of future HSI development with a proposal for using X-rays as a powerful technique to characterize magnetic materials. Using diffraction, small-angle scattering and reflectivity, the authors demonstrate the element sensitivity and strong dependence of the X-ray polarization on both the size and direction of the magnetic moments theoretically and experimentally.

The use of HSI goes beyond electromagnetic waves, with other available excitation sources such as X-rays and electrons, and new HSI modalities can also be extended to nanometer scales in spatial dimensions. This book brings together diverse HSI research areas to provide a comprehensive overview of the current status of machine- and deep-learning development for hyperspectral imaging.

> **Jung Y. Huang** Department of Photonics, Chiao Tung University, Hsinchu, Taiwan, Republic of China

Section 1

Overview of Hyperspectral

Imaging

### Section 1
