Preface

Glossary of Terms and Acronyms

Optical Remote Sensing

Spectral Coverage of Optical Remote Sensing

Spectral Characteristics of Earth Features

Spectral Resolution

Spatial Resolution

Pixel, Subpixel, and Superpixel

Radiometric Resolution

From Raw Data to Information Retrieval

Image Processing Techniques vs Image Types

Image Data Correction

Radiometric Errors Due to Atmosphere

Cloud Removal

Geometric Errors

Mapping Functions and Ground Control Points

Mapping Function Validation

Resampling

Image Registration Example

Image Radiometric Enhancement and Display

Image Histogram

Linear Histogram Modification

Linear Histogram Modification Example

Uniform Histogram and Cumulative Histogram

Histogram Equalization for Contrast Enhancement

Histogram Equalization Example

Color Composite Image Display

Principal Component Transformation for Image Display

Image Geometric Enhancement

Spatial Filtering

Image Smoothing

Speckle Removal

Edge and Discontinuity Detection

Spatial Gradient Detection

Morphological Operations

Hyperspectral Image Data Representation

Image Data Cube Files

Image Space and Spectral Space

Features and Feature Space

Pixel Vector, Image Matrix, and Data Set Tensor

Cluster Space

Image Clustering and Segmentation

Otsu’s Method

Clustering Using the Single-Pass Algorithm

Clustering Using the k-Means Algorithm

Clustering Using the k-Means Algorithm Example

Superpixel Generation Using SLIC

Pixel-Level Supervised Classification

Supervised Classification Procedure

Prototype Sample Selection

Training Samples and Testing Samples

Minimum Euclidean Distance Classifier

Spectral Angle Mapper

Spectral Information Divergence

Class Data Modeling with a Gaussian Distribution

Mean Vector and Covariance Matrix Estimation

Gaussian Maximum-Likelihood Classification

Other Distribution Models

Mahalanobis Distance and Classifier

k-Nearest Neighbor Classification

Support Vector Machines

Nonlinear Support Vector Machines

Handling Limited Numbers of Training Samples

Semi-Supervised Classification

Active Learning

Transfer Learning

Feature Reduction

The Need for Feature Reduction

Basic Band Selection

Mutual Information

Band Selection Based on Mutual Information

Band Selection Based on Class Separability

Knowledge-based Feature Extraction

Data-Driven Approach for Feature Extraction

Linear Discriminant Analysis

Orthogonal Subspace Projection

Adaptive Matched Filter

Band Grouping for Feature Extraction

Principal Components Transformation

Incorporation of Spatial Information in Pixel Classification

Spatial Texture Features using GLCM

Examples of Texture Features

Markov Random Field for Contextual Classification

Options for Spectral-Spatial-based Mapping

Subpixel Analysis

Spectral Unmixing

Endmember Extraction

Endmember Extraction with N-FINDR

Limitation of Linear Unmixing

Subpixel Mapping

Subpixel Mapping Example

Super-resolution Reconstruction

Artificial Neural Networks and Deep Learning with CNNs

Artificial Neural Networks: Structure

Artificial Neural Networks: Neurons

Limitation of Artificial Neural Networks

CNN Input Layer and Convolution Layer

CNN Padding and Stride

CNN Pooling Layer

CNN Multilayer and Output Layer

CNN Training

CNN for Multiple-Image Input

CNN for Hyperspectral Pixel Classification

CNN Training for Hyperspectral Pixel Classification

Multitemporal Earth Observation

Satellite Orbit Period

Coverage and Revisit Time

Change Detection

Classification Accuracy Assessment

Error Matrix for One-Class Mapping

Error Matrix for Multiple-Class Mapping

Kappa Coefficient Using the Error Matrix

Model Validation

Bibliography

Index

Hyper/multispectral imagery in optical remote sensing is an extension of color RGB pictures. The utilized wavelength range is beyond the visible, up to the reflective shortwave infrared. Hyperspectral imaging offers higher spectral resolution, leading to many wavebands. The spectral profiles recorded reveal reflected solar radiation from Earth-surface materials when the sensor is mounted on an airborne or spaceborne platform. An inverse process using machine-learning approaches is conducted for target detection, material identification, and associated environmental applications, which is the main purpose of remote sensing.

This Field Guide covers three areas: the fundamentals of remote sensing imaging for image understanding; image processing for correction and quality improvement; and image analysis for information extraction at subpixel, pixel, superpixel, and image levels, including feature mining and reduction. Basic concepts and fundamental understanding are emphasized to prepare the reader for exploring advanced methods.

I owe thanks to Professor John Richards, who introduced me to the remote sensing field and supervised my Ph.D. study on hyperspectral image classification. I learned a lot from John about critical thinking and thorough presentation.

I dedicate this Field Guide to my husband, Xichuan, my son, James, and my daughter, Jessica, who have shared and accompanied me on my learning journey in this subject for so many years and make me a proud career wife and mom.

Xiuping Jia
The University of New South Wales,
Canberra, Australia
May 2022