دانلود رایگان ترجمه مقاله ارزیابی اثربخشی روشهای درخت تصمیم برای طبقه بندی پوشش زمین (نشریه الزویر 2003)

ارزیابی اثربخشی روشهای درخت تصمیم برای طبقه بندی پوشش زمین

An assessment of the effectiveness of decision tree methods for land cover classification

Abstract

Choice of a classification algorithm is generally based upon a number of factors, among which are availability of software, ease of use, and performance, measured here by overall classification accuracy. The maximum likelihood (ML) procedure is, for many users, the algorithm of choice because of its ready availability and the fact that it does not require an extended training process. Artificial neural networks (ANNs) are now widely used by researchers, but their operational applications are hindered by the need for the user to specify the configuration of the network architecture and to provide values for a number of parameters, both of which affect performance. The ANN also requires an extended training phase.

In the past few years, the use of decision trees (DTs) to classify remotely sensed data has increased. Proponents of the method claim that it has a number of advantages over the ML and ANN algorithms. The DT is computationally fast, make no statistical assumptions, and can handle data that are represented on different measurement scales. Software to implement DTs is readily available over the Internet. Pruning of DTs can make them smaller and more easily interpretable, while the use of boosting techniques can improve performance.

In this study, separate test and training data sets from two different geographical areas and two different sensors—multispectral Landsat ETM+ and hyperspectral DAIS—are used to evaluate the performance of univariate and multivariate DTs for land cover classification. Factors considered are: the effects of variations in training data set size and of the dimensionality of the feature space, together with the impact of boosting, attribute selection measures, and pruning. The level of classification accuracy achieved by the DT is compared to results from back-propagating ANN and the ML classifiers. Our results indicate that the performance of the univariate DT is acceptably good in comparison with that of other classifiers, except with high-dimensional data. Classification accuracy increases linearly with training data set size to a limit of 300 pixels per class in this case. Multivariate DTs do not appear to perform better than univariate DTs. While boosting produces an increase in classification accuracy of between 3% and 6%, the use of attribute selection methods does not appear to be justified in terms of accuracy increases. However, neither the univariate DT nor the multivariate DT performed as well as the ANN or ML classifiers with high-dimensional data.

1 Introduction

The past three decades have seen continuing developments in the area of pattern recognition. Research into algorithmic aspects of pattern recognition has proceeded alongside the development of instruments that are capable of producing high volumes of data, including images with increasingly finer spatial and spectral resolution. After 30 years of satellite remote sensing of the Earth’s land surface, users of remotely sensed data now have access to sophisticated statistical and neural/connectionist algorithms for both fuzzy and hard classifications of their data (Mather, 1999; Schowengerdt, 1997).

Both the statistical and neural/connectionist approaches have limitations. Statistical methods rely on the assumption that the probabilities of class membership can be modelled by a specific probability density function. In most cases, the Gaussian distribution is chosen, as it is characterised by first- and second-order statistics, that is, the class mean vectors and class covariance matrices. If training set size is fixed, then the precision of the estimates of the elements of the sample class mean vector and sample class covariance matrix declines as the number of features (dimensions) increases, so that one might expect the performance of the classifier to degrade as the number of features increases. The assumption that the data in each class follow a multivariate normal model restricts the analysis to interval or ratio scale data.

Neural/connectionist methods appear to work well with training data sets that are smaller in size than those required for statistical procedures. On the other hand, network training times can be lengthy, while choice of the design of network architecture (in terms of numbers of hidden layers and neurons per layer) and the values of the learning rate parameters is not straightforward (Foody & Arora, 1997; Kavzoglu, 2001; Wilkinson, 1997). Unlike statistical methods, the neural/connectionist approach makes no assumptions concerning the statistical frequency distribution of the data or the measurement scales of the features that are used in the analysis. The most commonly used neural/ connectionist algorithm is the back-propagating multi-layer perceptron (Wilkinson, 1997), which is used in this study.

Decision tree (DT) classifiers have not been as widely used within the remote sensing community as either the statistical or the neural/connectionist methods. The advantages that decision trees offer include an ability to handle data measured on different scales, lack of any assumptions concerning the frequency distributions of the data in each of the classes, flexibility, and ability to handle non-linear relationships between features and classes (Friedl & Brodley, 1997). In contrast to neural networks, decision trees can be trained quickly, and are rapid in execution (Gahegan & West, 1998). They can be used for feature selection/reduction as well as for classification purposes (Borak & Strahler, 1999). Finally, the analyst can interpret a decision tree. It is not a ‘black box’, like the neural network, the hidden workings of which are concealed from view.

Overall classification accuracy is used here to measure the performance of the different methods. The level of classification accuracy that is achieved in a particular case depends on a number of factors, including the nature of the classification problem in terms of the complexity of the decision boundaries that separate the classes in feature space (assuming that the classes are separable), the training sample size, the adequacy of the training data in characterising the properties of the chosen classes, the dimensionality of the data, and the properties of the classifier used (Raudys and Pikelis, 1980). We do not consider all of these problems in this paper. However, the results of our analyses are internally comparable, as the same training and test data sets are used for all three classifiers, for two dissimilar study areas (Section 2). Thus, it is possible to examine both the relative performance of the different classifiers and the consistency of these comparisons between data sets with dissimilar characteristics in terms of the terrain of the study area and the nature of the imaging system used.

The paper is structured as follows. Section 2 describes the two test data sets that are used in this study. Brief details of the three classifiers are provided in Section 3. The effects of training set size, data dimensionality, attribute selection methods, pruning and boosting on the performance of the DT classifier are considered in Section 4. A short comparative analysis of the relative performance of the DT, artificial neural networks (ANNs), and maximum likelihood (ML) classifiers is given in Section 5, which is followed by a summary of conclusions.