ID: 53915
Title: Kernel-based quantitative remote sensing inversion
Author: Yanfei Wang, Changchun Yang and Xiaowen Li
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Quantitative remote sensing, Inversion
Abstract: To estimate structural parameters and spectral component signatures of Earth surface cover type, quantitative remote sensing seems to be an appropriate way of dealing with these problems. Since the real physical system that couples the atmosphere and the land surface is very complicated and should be continuous, sometimes it requires a comprehensive set of parameters to describe the system, so any practical physical model can only be approximated by a model that includes only a limited number of the most important parameters that capture the major variation of the real system.
The pivotal problem for quantitative remote sensing is the inversion. Inverse problems are typically ill-posed. The ill-posed nature is characterized by: (C1) the solution may not exist; (C2) the dimension of the solution space may be infinity; (C3) the solution is not continuous with the variation of the observed signals. These issues exist for all quantiative remote sensing inverse problems. For example, when sampling is poor, i.e., there are too few observations, or directions are poorly located, the inversion process would be underdetermined, which leads to the large condition number of the normalized system and significant noise propagation. Hence (C2) and (C3) would be the highlighted difficulties for quantitative remote sensing inversion.
This chapter will address the theory and methods from the viewpoint that the quantitative remote sensing inverse problems can be represented by kernel-based operator equations.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53914
Title: Kernel methods for unmixing hyperspectral imagery
Author: Joshua Broadwater, Amit Banerjee and Philippe Burlina
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Linear mixtures, Non linear mixtures
Abstract: This chapter introduces the concept of kernel unmixing algorithms for hyperspectral imagery. The physics of linear and nonlinear mixtures are explored and shown to be suited for kernel approximations. Based on this evidence, a fully automatic kernel unmixing algorithm is developed for endmember extraction and abundance estimation. This kernel unmixing framework has the flexibility to account for both linear and nonlinear behaviour by using different kernel functions. Additionally to support nonlinear mixtures, a new physics-based kernel is proposed based on the physics of intimate mixtures. Experimental results on both in-lab and real-world hyperspectral data show both qualitative and quantitative improvements when using the proposed kernel unmixing approach.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53913
Title: Mean kernels for semi-supervised remote sensing image classification
Author: Mattia Marconcini and Lorenzo Bruzzone and Gustavo Camps-Valls
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: semi-supervised support vector machine (SVM), expectation-maximization (EM), Gaussian mixture models (GMM)
Abstract: This chapter presents a semi-supervised support vector machine (SVM) classifier based on the combination of the expectation-maximization (EM) algorithm for Gaussian mixture models (GMM) and the mean map kernel. The proposed method improves classification accuracy by reinforcing samples in the same cluster belonging to the same class through the use of composite kernels learned from the testing image. These sample and cluster similarities are included in the standard SVM by means of a linear combination of kernels. The cluster similarity is directly computed in the kernel space with a dedicated kernel that is based on the means of the feature vectors in this space. Moreover, only the most reliable samples in terms of likelihood values are used to compute a kernel function that accurately reflects the similarity between clusters in the feature space. Results show that the proposed approach is especially suited for situations where the available labelled information does not properly describe the classes in the test image.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53912
Title: A domain adaptation SVM and a circular validation strategy for land-cover maps updating
Author: Mattia Marconcini and Lorenzo Bruzzone
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: SVMs (DASVM), land-cover maps
Abstract: In this paper, we present a novel domain adaptation classifier based on SVMs (DASVM) for land-cover maps updating, which can be employed in real operational situations when ground-truth labels are available only for a reference image acquired over the investigated geographical area before the one to classify. In addition, we also propose a circular validation strategy for the accuracy assessment of the classification results when no labelled samples are available for the image to be categorized.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53911
Title: One-class SVMs for hyperspectral anomaly detection
Author: Amit Banerjee, Philippe Burlina and Chris Diehl
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Hyperspectral Support Vector Data Description (SVDD) ,Reed-Xiaoli (RX) ,constant false-alarm rate (CFAR) ,receiver operating characteristic (ROC) ,Wide Area Airborne Mine Detection (WAAMD) ,Digital Imagery Collection Experiment (HYDICE)
Abstract: A support vector framework for hyperspectral anomaly detection is developed in this chapter. Conventional methods for detecting anomalies in hyperspectral images are based on the popular Reed-Xiaoli (RX) detector. However, these algorithms typically suffer from a large numbers of false alarms, due to the assumptions that the background is Gaussian and homogeneous. In practice, these assumptions are often violated, especially when the neighbourhood of a pixel contains multiple types of terrain.
To remove these assumptions, a novel anomaly detector is proposed and derived that incorporates a nonparametric background model based on the Support Vector Data Description (SVDD). The SVDD is a one-class support vector classifier that can model the support of a distribution. Expanding on prior work, a geometric interpretation of the SVDD is developed to propose a decision rule that utilizes a new test statistic and shares some of the properties of constant false-alarm rate (CFAR) detectors. Two versions of the algorithm are presented to detect either local or global anomalies and an analysis of their computational requirements is provided. Using curves, the improved performance and reduction in the false alarm rate when using the SVDD-based detector are demonstrated on Wide Area Airborne Mine Detection (WAAMD) and Hyperspectral Digital Imagery Collection Experiment (HYDICE) hyperspectral imagery.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53910
Title: Target detection with kernels
Author: Nasser M. Nasrabadi
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Principal Component Analysis (PCA), Fisher Linear Discriminant (FLD) Analysis, the Eigenspace Separation Transform (EST), Reed-Xiaoli (RX), receiver operating characteristics (ROC)
Abstract: This book chapter provides a performance comparison of various linear and nonlinear anomaly detection techniques based on kernels. Three different subspace anomaly detectors based on Principal Component Analysis (PCA), Fisher Linear Discriminant (FLD) Analysis, and the Eigenspace Separation Transform (EST) are reviewed and their nonlinear versions (kernel extensions) are discussed. In addition to the subspace-based anomaly detectors, the well known Reed-Xiaoli (RX) anomaly detector and its kernel version are also implemented. Comparisons between all linear and nonlinear anomaly detectors are made using receiver operating characteristics (ROC) curves for simulated toy data and real hyperspectral imagery.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53909
Title: Multi-temporal image classification with kernels
Author: Jordi Munoz-Mari, Luis Gomez-Chova, Manel Martinez-Ramon, Jose Luis Rojo-Alvarez, Javier Calpe-Maravilla and Gustavo Camps-Valls
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Support Vector Machine classifier (SVM),Support Vector Domain Description (SVDD) classifier, multi-sensor images, nonlinear kernel classifiers
Abstract: In generic problems of multi-temporal classification of remote sensing images, different sources of information such as temporal, contextual or multi-sensor, are commonly available. The combination of these heterogenous sources of information is still an active research area. In this chapter, we present a classification framework based on kernel methods for multitemporal classification of remote sensing images. The proposed kernel classifiers not only process multi-temporal images simultaneously, and with different levels of sophistication, but also allow one to properly combine different data sources, such as contextual information and multi-sensor images. In this chapter, we also present two nonlinear kernel classifiers for well-known change detection methods formulating them in an adequate high dimensional kernel-induced feature space. The developed kernels are used in two core classification machines: the binary Support Vector Machine classifier (SVM), and the one-class Support Vector Domain Description (SVDD) classifier.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53908
Title: Kernel Fisher ' s Discriminant with heterogeneous kernels
Author: M. Murat Dundar and Glenn Fung
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Fisher ' s Discriminant (KFD), heterogeneous kernel models (AKFD), Hymap dataset
Abstract: In this chapter we first present a framework suitable for obtaining a nonlinear version of the Fisher ' s Discriminant (KFD). Then we propose an iterative classification algorithm for KFD using heterogeneous kernel models (AKFD). In contrast with the standard KFD that requires the user to predefine a kernel function, we incorporate the task of choosing an appropriate kernel into the optimization problem to be solved. The choice of kernel is defined as a linear combination of kernels belonging to a potentially large family of different positive semi-definite kernels. Experiments on a Hymap dataset demonstrate that the AKFD algorithm outperforms the linear version of the Fisher ' s discriminant and also significantly reduces the time required to train the KFD algorithm while maintaining similar performance.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53907
Title: On training and evaluation of SVM for remote sensing applications
Author: Giles M. Foody
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Image classification, SVM
Abstract: The design of the training and testing stages of a supervised classification is, to differing degrees, classifier-dependent. This chapter provides an overview of the some of the key issues in the design of training and testing stages for image classification, with particular regard to classification by SVM. A key issue stressed is that only effective support vectors are required in training an SVM. This feature enables SVM to derive accurate classifications from small training sets. The accuracy of an SVM may be assessed using a variety of approaches but some may not be practical for all types of SVM classification.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53906
Title: The Support Vector Machine (SVM) algorithm for supervised classification of hyperspectral remote sensing data
Author: J. Anthony Gualtieri
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Support Vector Machine (SVM), Hyperspectral remote sensing data, non parameteric SVM
Abstract: The Support Vector Machine (SVM) algorithm provides an effective way to perform supervised classification of hyperspectral remote sensing data. The problem is to learn from a training set of examples-hyperspectral data with class labels attached- and then generalize to find the class labels of hyperspectral data outside the training set. The high dimensionality of hyperspectral data, due to the many spectral channels is a sensor simultaneously measures, causes problems for other supervised classification algorithms, both parametric and nonparametric. A key feature of the nonparameteric SVM superivsed classification algorithm is its ability to use high-dimensional data without the usual recourse to a feature selection step to reduce the dimensionality of the data, as is required by many other algorithms in order for them to work. Additionally, SVM can utilize the Kernel method of projecting the data into a different space to improve performance. We give an introduction to hyperspectral data and its acquisition and for most of the sequel focus on using only the spectral information available in the data for performing the classification. We then present an overview of the mathematical foundations of statistical learning theory, show how the large margin SVM, appropriate to supervised classification, can be derived in the context of these very general results, show its realization as a quadratic optimization problem, and indicate the Kernel method, which extends the efficacy of the SVM by using nonlinear transformation of the training data. These results are then applied to several benchmark hyperspectral data sets, and the SVM results are compared with other supervised classification methods. Then we indicate how using the spatial content of the data can furthr improve the classification results. Finally we close with an exploration of the reasons why the SVM offers improved performance over other algorithms and summarize with a brief conclusion.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53905
Title: An introduction to kernel learning algorithms
Author: Peter V. Gehler and Bernhard Scholkopf
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons, Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Kernel learning algorithms, Gaussian Processes, kernel principal analysis
Abstract: Kernel learning algorithms are currently becoming a standard tool in the area of machine learning and pattern recognition. In this chapter we review the fundamental theory of kernel learning. As the basis building block we introduce the kernel function, which provides an elegant and general way to compare possibly very complex objects. We then review the concept of a reproducing kernel Hilbert space and state the representer theorem. Finally we give an overview of the most prominent algorithms, which are support vector classification and regression, Gaussian Processes and kernel principal analysis. With multiple kernel learning and structured output prediction we also introduce some more recent advancements in the field.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53904
Title: Machine learning techniques in remote sensing data analysis
Author: Bjorn Waske, Mathieu Fauvel, Jon Atli Benediktsson and Jocelyn Chanussot
Editor: Gustavo Camps-Valls, Lorenzo Bruzzone
Year: 2009
Publisher: John Wiley and Sons Ltd, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Kernel Methods for Remote Sensing Data Analysis
Keywords: Machine learning, Neural networks, Support vector machines, Multiple classifier systems
Abstract: Several applications have been developed in the field of remote sensing image analysis during the last decades. Besides well-known statistical approaches, many recent methods are based on techniques taken from the field of machine learning. A major aim of machine learning algorithms in remote sensing is supervised classification, which is perhaps the most widely used image classification approach. In this chapter a brief introduction to machine learning and the different paradigms in remote sensing is given. Moreover this chapter briefly discusses the use of recent developments in supervised classification techniques such as neural networks, support vector machines and multiple classifier systems.
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53903
Title: Freshwater Ecology -Concepts & Environment Applications of Limnology
Author: Walter K Dodds and Matt R. Whiles
Editor: Walter K Dodds and Matt R. Whiles
Year: 2010
Publisher: Academic Press, Second Edition, 2010
Source: Centre for Ecological Sciences
Reference: None
Subject: Freshwater Ecology -Concepts & Environment Applications of Limnology
Keywords: None
Abstract: None
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53902
Title: Functional assessment of Wetlands Towards evaluation of ecosystem services
Author: Edward Maltby
Editor: Edward Maltby
Year: 2009
Publisher: Woodhead publishing Limited & CRC Press LLC, 2009
Source: Centre for Ecological Sciences
Reference: None
Subject: Functional assessment of Wetlands Towards evaluation of ecosystem services
Keywords: None
Abstract: None
Location: 215
Literature cited 1: None
Literature cited 2: None
ID: 53901
Title: Ground Object Identification-based on Absorption-band Position using EO-1 Hyperion Data
Author: Xu Yuanjin . Zhang Zhenfei . Hu Guangdao
Editor: Prof B.L.Deekshatulu
Year: 2010
Publisher: Indian Society of Remote Sensing, Vol 38, No 2, June 2010
Source: Centre for Ecological Sciences
Reference: None
Subject: Journal of the Indian Society of Remote Sensing
Keywords: Hyperspectral imagery . Ground object identification . Absorption-band position
Abstract: In order to accurately identify ground
objects in the hyperspectral imagery by spectral
matching, it is important to analyze the absorptionband
parameters. This paper presents a new spectral
matching method which is based mainly on analysis
of the absorption-band position. A measured
spectrum of a ground object can be subject to shifts
from its real wavelength position; meanwhile an
absorption band in the spectrum can also be shifted
relatively. Both these shifts are due to theenvironmental effects. Our spectral matching method
stresses the quantification of the total shift of the
absorption-band position, thus to get a possible offset
range of the measured absorption bands. This offset
range is taken as a constraint on the matching process.
The pixel spectrum in the image is then compared to
each known reference spectrum in a spectral library
previously built, so that the ground object
corresponding to the reference spectrum is identified.
A case study is conducted in Pulang Porphyry Copper
deposit, Zhongdian county, Yunnan, China. Five
types of ground objects were studied and it is shown
that our methods can get more accurate identification
results than the approach which does not consider
the shift ranges.
Location: 215
Literature cited 1: None
Literature cited 2: None