Factors influencing the accuracy of remote sensing classifications: a comparative study

Pal, Mahesh (2002) Factors influencing the accuracy of remote sensing classifications: a comparative study. PhD thesis, University of Nottingham.

[img]
Preview
PDF
4Mb

Abstract

Within last 20 years, a number of methods have been employed for classifying remote sensing data, including parametric methods (e.g. the maximum likelihood classifier) and non-parametric classifiers (such as neural network classifiers).Each of these classification algorithms has some specific problems which limits its use. This research studies some alternative classification methods for land cover classification and compares their performance with the well established classification methods. The areas selected for this study are located near Littleport (Ely), in East Anglia, UK and in La Mancha region of Spain. Images in the optical bands of the Landsat ETM+ for year 2000 and InSAR data from May to September of 1996 for UK area, DAIS hyperspectral data and Landsat ETM+ for year 2000 for Spain area are used for this study. In addition, field data for the year 1996 were collected from farmers and for year 2000 were collected by field visits to both areas in the UK and Spain to generate the ground reference data set.

The research was carried out in three main stages.The overall aim of this study is to assess the relative performance of four approaches to classification in remote sensing - the maximum likelihood, artificial neural net, decision tree and support vector machine methods and to examine factors which affect their performance in term of overall classification accuracy.

Firstly, this research studies the behaviour of decision tree and support vector machine classifiers for land cover classification using ETM+ (UK) data. This stage discusses some factors affecting classification accuracy of a decision tree classifier, and also compares the performance of the decision tree with that of the maximum likelihood and neural network classifiers. The use of SVM requires the user to set the values of some parameters, such as type of kernel, kernel parameters, and multi-class methods as these parameters can significantly affect the accuracy of the resulting classification. This stage involves studying the effects of varying the various user defined parameters and noting their effect on classification accuracy. It is concluded that SVM perform far better than decision tree, maximum likelihood and neural network classifiers for this type of study.

The second stage involves applying the decision tree, maximum likelihood and neural network classifiers to InSAR coherence and intensity data and evaluating the utility of this type of data for land cover classification studies. Finally, the last stage involves studying the response of SVMs, decision trees, maximum likelihood and neural classifier to different training data sizes, number of features, sampling plan, and the scale of the data used. The conclusion from the experiments presented in this stage is that the SVMs are unaffected by the Hughes phenomenon, and perform far better than the other classifiers in all cases. The performance of decision tree classifier based feature selection is found to be quite good in comparison with MNF transform. This study indicates that good classification performance depends on various parameters such as data type, scale of data, training sample size and type of classification method employed.

Item Type:Thesis (PhD)
Supervisors:Mather, Paul
Uncontrolled Keywords:land cover classification, support vector machines, decision tree classifier
Faculties/Schools:UK Campuses > Faculty of Social Sciences, Law and Education > School of Geography
ID Code:314
Deposited By:Mahesh Pal
Deposited On:20 Mar 2008
Last Modified:06 Feb 2009 14:43

Repository Staff Only: item control page