Implement Of Decision Tree Classifier And Application In Remote Sensing Image Classification

As one of the main classification methods in Data Mining, the Decision Tree algorithmdoes not need any prior knowledge or preferences and is suitable for exploratory knowledgediscovery. With clearly defined structures, fast operation speed, high accuracy, flexibility androbustness, the Decision Tree algorithm can be used to handle high-dimensional data and theknowledge acquired is intuitive and easy to be understood. Now the Decision Tree algorithmhas been widely used in medicine, manufacturing and production, financial analysis,astronomy, molecular biology and remote sensing image classification, etc.Remote sensing image classification is one of the main means of remote sensing imageinterpretation and its basic idea is to select features based on the analysis of spectral, spatial,geometry, texture in remote sensing image, then classify every pixel by some certain means.The methods of remote sensing image classification can be divided into pixel-based andobject-oriented. For reducing the manual workload, Nearest Neighbor, Maximum LikelihoodClassification, Support Vector Machine (SVM), Fuzzy Clustering, Decision Tree and NeuralNetwork et al algorithms have been used in the pixel-based classification of remote sensingimage as well as Nearest Neighbor, Membearship Function, Support Vector Machine andDecision Tree et al algorithms have been used in object-oriented classification of remotesensing image. The Decision Tree algorithm does better than other classification algorithmswith remote sensing image that has feature spatial distribution is complex. Further more thedecision tree or its rules can be analyzed and modified by experts and input to expert system.Now the Decision Tree algorithm has been used in remote sensing image classificationin some studies. But in these studies, the building of Decision Tree depends on existing datamining software and few deep research work focus on decision tree algorithm, also lacks thesoftware of remote sensing classification based on Decision Tree algorithm. Based onBoostTree algorithm, this paper proposed a new algorithm of decision tree ensembles forremote sensing image classification–AdaTree.WL. With AdaTree.WL algorithm, a piece ofsoftware was developed for pixel-based and object-oriented remote sensing imageclassification. In the experiment, Landsat7ETM+images are classified by pixel-basedmethod and Worldview-2images are classified by object-oriented method. The AdaTree.WLalgorithm is compared with other classifiers. The main research contents and results show asfollows:Firstly, based on the study of Decision Tree algorithm and remote sensing imageclassification, the AdaTree.WL algorithm is proposed in this paper by using BoostTree astemplate. It is a combination of C4.5and AdaBoost.M1algorithms. In the AdaTree.WLalgorithm, the structure of C4.5and the final hypothesis of AdaBoost.M1are modified. In this paper，a classifier is designed with AdaTree.WL algorithm, and the classifier is named GLCtree.Secondly, based on the above study, according to analysis and summary of remotesensing image classification, a method of using AdaTree.WL algorithm as the classifier inpixel-based and object-oriented remote sensing image classification is found. With themethod, a piece of software is designed. The software can not only be used in pixel-basedremote sensing image classification, but also object-oriented remote sensing imageclassification with the segmentation of remote sensing image and it solves the problem ofdepending on existing Data Mini software when using Decision Tree algorithm to classifyremote sensing image.Thirdly, through conducting the experiment of pixel-based classification using Landsat7ETM+images and the experiment of object-oriented classification using WorldView-2images, the AdaTree.WL algorithm is compared with other classification algorithms such asBoostTree, C5.0and Support Vector Machine algorithms. The results show that, theAdaTree.WL algorithm gets same accuracy compared with C5.0algorithm and greateraccuracy than BoostTree and Vector Machine algorithms. Its average Kappa coefficientsreaches0.9052in pixel-based classification using Landsat7ETM+images and0.9398inobject-oriented classification using WorldView-2images. Also, the features can be filtered bycalculating the contribution value depending on AdaTree.WL algorithm.