Gustavo Camps-Valls on Kernel-Based Methods for Hyperspectral Image Classification
Fast Moving Front Commentary, May 2011
|
Article: Kernel-based methods for hyperspectral image classification
Authors: Camps-Valls, G;Bruzzone, L |
Gustavo Camps-Valls talks with ScienceWatch.com and answers a few questions about this month's Fast Moving Fronts paper in the field of Engineering.
Why do you think your paper is highly
cited?
The paper presented an exhaustive comparison of methods for remote sensing image classification. The family of methods collectively known as "kernel methods" was somewhat new for the remote sensing scientific community at that time. We showed both theoretically and experimentally the powerfulness of these family of methods, and explained their theoretical relations quite intuitively. The results were compared to standard methods used so far in a very detailed way. After this work, many researchers in the field turned to using kernel machines for their applications and assessed the usefulness themselves.
Does it describe a new discovery, methodology, or
synthesis of knowledge?
"By monitoring urban growth, estimating temperature, moisture or ocean salinity, and identifying objects on the surface, these methods provide valuable information for policy and decision makers, as well as for tourism, defense or climage change applications."
It is mainly a synthesis of previous knowledge: the methods already existed but were not exploited in the field of remote sensing data analysis in a systematic way. We provided a theoretical overview of the methods and an extensive experimental validation to stress the particularly useful properties for this kind of data.
Would you summarize the significance of your paper
in layman's terms?
The paper proposes mathematical methods to efficiently process the images acquired by the satellite sensors. With these tools, international organizations like ESA, NASA, or EUMETSAT can give citizens better weather predictions, identify changes in the land production, detect particular materials on the Earth's surface, monitor climate change, prevent and manage natural disasters, etc.
How did you become involved in this research, and
how would you describe the particular challenges, setbacks, and
successes that you've encountered along the way?
The paper was done in collaboration with Dr. Lorenzo Bruzzone at the University of Trento (Italy). The research was conducted during my stay at his lab in a very hot summer time. I enjoyed a lot learning and collaborating with him and his team. I must say that I also learned Italian and discovered that beautiful region. Since then we have collaborated quite a lot, and gotten involved in common research projects in which we shared students for short periods.
The initial idea of comparing classification tools turned to be more complex and articulated after a while, also because we aimed at studying the theoretical features of the methods with regard to the statistical properties of the remote sensing data. The preliminary results we obtained were somewhat expected.
We extended the work by including more theoretical analysis and experimental results in different (potentially interesting) situations: we wanted to see the performance of the methods in some extreme situations that the final user would face: very high noise levels, few training examples, high dimensions, etc. After some months, we aimed at fixing some misconceptions we saw in the literature, and to do a seminal work in what we believed it would become a common toolbox for users and researchers in the field.
The work was conducted thanks to the funding I got from my university. Many methods were developed in-home and others kindly delivered by other researchers: SVM code was kindly provided by J. Ma, and Gunnar Raetsch shared with us his nice implementation of AdaBoost. Also, the need for representative datasets was an issue. At this point, we should thank Dr. David Landgrebe for providing the hyperspectral AVIRIS data. I should also thank my colleague and friend Dr. José Luis Rojo for the useful comments and suggestions on the paper.
Where do you see your research leading in the
future?
The methods we inspected here have been evolving in the last few years in different ways. On the one hand, people are studying how to incorporate prior information on kernels. On the other hand, adapting the formulation to the particular problem is a trending topic: nowadays kernel methods are useful not only for classification but also for feature extraction, regression and function approximation, density estimation, or target detection.
In the future I foresee more theoretical insight in the methods, e.g., studying the particular metrics induced by arbitrary implicit mappings, the pre-image issue, and how to extract knowledge from built machines.
Do you foresee any social or political
implications for your research?
Sure, we think that the methods have a great societal value. For instance
urban monitoring, fire detection, or flood prediction from remotely sensed
multispectral or radar images have a great impact on economical and
environmental issues. By monitoring urban growth, estimating temperature,
moisture or ocean salinity, and identifying objects on the surface, these
methods provide valuable information for policy and decision makers, as
well as for tourism, defense, or climate change
applications.
Gustavo Camps-Valls, PhD in Physics
Image Processing Laboratory (IPL)
Universitat de València
València, Spain
Web |
Web
KEYWORDS: ADABOOST, FEATURE SPACE, HYPERSPECTRAL CLASSIFICATION, KERNEL-BASED METHODS, KERNEL FISHER DISCRIMINANT ANALYSIS, RADIAL BASIS FUNCTION NEURAL NETWORKS, REGULARIZATION, SUPPORT VECTOR MACHINES, LAND-COVER CLASSIFICATION, REMOTE-SENSING IMAGES, NEURAL NETWORKS, CLASSIFIERS, ROBUST.