The Role of Hyperspectral Imaging: A Literature Review

Muhammad Mateen, Junhao Wen, Nasrullah Nasrullah, Muhammad AzeemInternational Journal of Advanced Computer Science and Applications 2018 · DOI

Abstract Optical analysis techniques are used recently to detect and identify objects from a large scale of images. Hyperspectral imaging (HSI) technique is one such method. While human vision is based on three primary color bands (red, green, and blue), spectral imaging divides vision into many more bands. Hyperspectral remote sensors acquire imagery data in the form of hundreds of adjoining spectral bands. This paper illustrates fundamental concepts, remote sensing information, HSI methods, and applications. Furthermore, in a forensic context, novel methods involving deep neural networks are elaborated, offering a foundation for future research in HSI using deep learning.

Introduction: Beyond Human Vision

While the human eye and standard cameras perceive the world through three primary color channels—Red, Green, and Blue—many materials possess unique "spectral signatures" that are only discernible in much finer slices of the electromagnetic spectrum. Hyperspectral Imaging (HSI) captures hundreds of these narrow, adjacent bands, creating a high-dimensional "fingerprint" for every pixel.

The primary challenge in HSI is the sheer volume of data. These datasets are often noisy and computationally expensive to process. This review addresses the transition from traditional image processing to sophisticated spectral analysis, specifically highlighting a critical gap in forensic science. Traditional methods for locating buried remains, such as search dogs or ground-penetrating radar, are often slow or ineffective in challenging terrains. HSI emerges as a powerful remote sensing alternative, capable of detecting subtle chemical and biological "halos" resulting from decomposition.

Methodology: Spectral vs. Image Analysis

The paper categorizes HSI methodology into two primary workflows: Spectral Analysis and Image Analysis.

The dual workflow of hyperspectral imaging: Spectral analysis focuses on extracting detailed signatures, while image analysis optimizes for rapid computation through band selection.
  • Spectral Analysis: This approach focuses on the "peaks and valleys" of light reflectance. It involves spectral extraction, preprocessing, and model calibration using statistical methods like Principal Component Analysis (PCA) for dimensionality reduction and Partial Least Squares (PLS) for regression.
  • Image Analysis: This workflow selects specific, important wavebands to enable rapid computation. It includes feature extraction and the development of calibrated models for specific tasks.
  • Deep Learning Integration: There is a significant shift toward Deep Neural Networks (DNNs), including Auto-Encoders and Convolutional Neural Networks (CNNs). A notable advancement is the R-VCANet model, which combines a "Rolling Guidance Filter" with a "Vertex Component Analysis Network," proving effective even with limited labeled training data.

Applications and Key Insights

The review synthesizes findings across diverse fields, from agriculture to medicine.

  • The Vegetation Proxy: A compelling forensic insight is that HSI can detect buried remains indirectly. Decomposing bodies alter soil chemistry, which subsequently affects the pigments and health of local vegetation. This "stress response" creates a detectable spectral signature from the air.
  • The Curse of Dimensionality: In HSI, more data is not always better. Redundancy across hundreds of bands can introduce noise. Effective "band selection" is critical to streamline processing without losing vital information.
  • Medical and Food Safety: HSI has demonstrated exceptional precision, achieving a 99% accuracy rate in classifying meat marbling and a 100% success rate in detecting poultry egg fertility at "Day 0," before any visible changes occur.

Forensic Detection and Deep Learning Results

In the forensic domain, the integration of deep learning has revolutionized the identification of clandestine graves. By processing HSI "patches" through a hybrid feature extractor, researchers can identify burial sites based on gas emissions (such as CH4, N2O, and CO2) and vegetation stress.

A Deep Neural Network architecture designed to classify hyperspectral image patches, providing a binary output to indicate the presence of buried remains.

Deep learning models consistently outperform traditional linear models in these complex environments. They can reduce high-dimensional HSI data into a simple binary output (1 or 0), indicating the presence or absence of buried remains with higher precision than manual feature extraction.

Limitations and Future Outlook

Despite its potential, HSI faces significant hurdles in computational complexity. Real-time processing of hundreds of spectral bands requires substantial hardware power, often necessitating high-end GPUs. Furthermore, HSI is sensitive to atmospheric conditions and sensor noise, often requiring manual band removal.

Future research is directed toward the full automation of forensic searches. The goal is to deploy Unmanned Airborne Vehicles (UAVs) that autonomously scan vast areas and utilize deep learning to generate "prediction maps." Such systems would provide investigators with high-interest sites, drastically reducing the time and human cost associated with search and recovery operations.

Summary updated

Abstract

Optical analysis techniques are used recently to detect and identify the objects from a large scale of images. Hyperspectral imaging technique is also one of them. Vision of human eye is based on three basic color (red, green and blue) bands, but spectral imaging divides the vision into many more bands. Hyperspectral remote sensors achieve imagery data in the form of hundreds of adjoining spectral bands. In this paper, our purpose is to illustrate the fundamental concept, hyperspectral remote sensing, remotely sensed information, methods for hyperspectral imaging and applications based on hyperspectral imaging. Moreover, in the forensic context, the novel methods involving deep neural networks are elaborated in this paper. The proposed idea can be useful for further research in the field of hyperspectral imaging using deep learning.
BibTeX
@article{mateen2018role,
  author = {Muhammad Mateen and Junhao Wen and Nasrullah Nasrullah and Muhammad Azeem},
  title = {{The Role of Hyperspectral Imaging: A Literature Review}},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2018},
  doi = {10.14569/ijacsa.2018.090808},
}
Figure