STUDY AND IMPLEMENTATION OF SOME DECISION TREE ALGORITHMS FOR IMAGE CLASSIFICATION

Abstract

In recent years, machine learning and data mining technique have become increasingly common in remote sensing applications. One area in which such techniques are particularly useful is classification of remotely sensed data for mapping applications. Advancement in computation tools have resulted in more and more information extraction from remotely sensed data. Development of new method and algorithms shall always be required to improve the quality of information extraction. Over the years, a number of computationally efficient machine learning algorithms have been developed these artificial neural network (ANN) algorithms, support vector machine (SVM), evidential reasoning, decision tree classifier, and many others. Choice of a classification algorithm is generally based upon a number of factors, among which are availability of software, ease of use, and performance in terms of classification accuracy. However, each has its own limitations. An effective algorithm for image classification should be computationally efficient. In addition, it should be robust to handle noisy data, errors, and redundant information's in the datasets. Decision tree classification techniques have been used for a wide range of classification of remotely sensed data. It has its advantages in image classification due to there flexibility, nonparametric nature, and ability to handle non-linear relations between features and classes. Various algorithms have been proposed for the generation of decision trees. The main objective of this dissertation is to examine the efficacy of four decision tree classification algorithm namely Gini index, Towing rule, Chi-square statistic, Information gain. Decision tree induction methods such as, CART and ID3 are used implement these algorithm using Visual C++ language. The classification performances. of decision trees classifier have been compared with conventional maximum likelihood classifier. Since these algorithms are not available in any commercial remote sensing image processing packages, so suitable computer code have been developed and implemented for the present study. The traditional parametric classifiers is require training data to be normally distributed, due to this fact it becomes difficult to add ancillary layers into classification procedures to improve accuracy of the classification. Decision tree is an overcome to this II type of problem because its capability of multi-source classification. In the second experiment of this dissertation the ancillary data i.e. DEM, slope and aspect are used to remove the spectral confusion of dataset and enhance the classification accuracy. This information's can be used in categorized form also in decision tree classification. Decision tree classifiers can be used these information to enhance the classification accuracy but maximum likelihood responds in les accuracy. CART based algorithm mostly Gini index perform relatively best in this experiment for adding the ancillary layers. Multi-source classification can use in ETM+ dataset where 6 spectral bands are beneficial to ID3 algorithm of decision tree to enhance the accuracy. In supervised classification of land cover using high dimensional remote sensing images decision tree classifier is beneficial. On classifying a hyper-spectral dataset decision tree classifier demonstrates its efficiency over that. Maximum likelihood classifier perform an accuracy of 44.72 % as the classifying dataset in 10 classes and in decision tree classifiers information gain produces the highest accuracy of 73.68%. This experiment also shows the capability decision tree classifier of handling the small training dataset to perform classification accuracy, on other side maximum likelihood classifier bounded by its limitations. In all experiments in this dissertation the efficacy of Decision Tree classifiers shows its use in training and classification of remotely sensed image data. Due to its. capability of generating human interpretable decision rules in relatively fast speed it can be used as the alternative for the knowledge base classifiers and the expert systems. III