On the extraction and representation of land cover information derived from remotely sensed imagery
On the extraction and representation of land cover information derived from remotely sensed imagery
The main contributions of this thesis consist of a novel probabilistic interpretation of subpixel area proportions that has a number of important implications: It is used to motivate a new probabilistic notation for area proportion information that, due to the probabilistic interpretation, is simple and intuitive, to show that certain types of fuzzy classifier have an equivalent interpretation as crisp classifiers, a relation that can be used to prove that they are capable of producing optimal proportion estimates and which suggests a number of enhancements that are shown empirically to improve the fuzzy classifiers performance. Finally, the probabilistic interpretation is used to provide insights into the application of the cross entropy error function in fuzzy classification that are shown to be supported by empirical evidence.
The thesis also presents a novel analysis of the impact of the sensor point spread function on fuzzy classifier performance that shows that the problem of extracting subpixel proportion information from pixels' spectral signatures is ill-posed. This is used to motivate the use of a new representation for subpixel proportion information - the spectrum conditional proportion distribution - that overcomes many of the limitations of the standard representation. Specifically, the distribution can fully represent the proportion information in a pixel's spectral signature, it permits this information to be propagated without loss, and it allows different sources of proportion information to be optimally combined. A number of techniques for extracting proportion distributions are described and empirical results are presented that underlie the utility of the new representation.
University of Southampton
Manslow, John
ec408a28-1e96-4fd6-acd7-b54ef3c6dbef
2001
Manslow, John
ec408a28-1e96-4fd6-acd7-b54ef3c6dbef
Manslow, John
(2001)
On the extraction and representation of land cover information derived from remotely sensed imagery.
University of Southampton, Doctoral Thesis.
Record type:
Thesis
(Doctoral)
Abstract
The main contributions of this thesis consist of a novel probabilistic interpretation of subpixel area proportions that has a number of important implications: It is used to motivate a new probabilistic notation for area proportion information that, due to the probabilistic interpretation, is simple and intuitive, to show that certain types of fuzzy classifier have an equivalent interpretation as crisp classifiers, a relation that can be used to prove that they are capable of producing optimal proportion estimates and which suggests a number of enhancements that are shown empirically to improve the fuzzy classifiers performance. Finally, the probabilistic interpretation is used to provide insights into the application of the cross entropy error function in fuzzy classification that are shown to be supported by empirical evidence.
The thesis also presents a novel analysis of the impact of the sensor point spread function on fuzzy classifier performance that shows that the problem of extracting subpixel proportion information from pixels' spectral signatures is ill-posed. This is used to motivate the use of a new representation for subpixel proportion information - the spectrum conditional proportion distribution - that overcomes many of the limitations of the standard representation. Specifically, the distribution can fully represent the proportion information in a pixel's spectral signature, it permits this information to be propagated without loss, and it allows different sources of proportion information to be optimally combined. A number of techniques for extracting proportion distributions are described and empirical results are presented that underlie the utility of the new representation.
Text
786751.pdf
- Version of Record
More information
Published date: 2001
Identifiers
Local EPrints ID: 464367
URI: http://eprints.soton.ac.uk/id/eprint/464367
PURE UUID: 5fd57b9e-0111-453c-a4f7-e12f04eb53a4
Catalogue record
Date deposited: 04 Jul 2022 22:21
Last modified: 16 Mar 2024 19:27
Export record
Contributors
Author:
John Manslow
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics