Feature Extraction Of Typical Vegetation Based On RapidEye Images

Vegetation is an important object in the geographic conditions monitoring, it has a profoundimpact on the environmental quality of human ecology. As the vegetation types complex anddiverse in nature, the traditional site survey using artificial methods will spends a lot ofmanpower and material resources. In recent years, remote sensing technology for theclassification and identification of vegetation provides a new way. Remote Sensing image canrecord the real vegetation and environment information comprehensively, and because ofdifferent vegetation types has different spectral characteristics, it makes possible to distinguishthem. With the improvement of remote sensing technology, high-resolution images can providemore features about spatial, geometry and texture, which will improve the interpretationaccuracy.Aiming at high-resolution remote sensing image RapidEye has abundant texture, structureand spatial features, this paper combined with multi temporal and multi sensors, using differentclassification methods to extract diffirent vegetation types. The main research contents show asfollows:(1) Through improving the binding mode between AdaBoost algorithm and the decision treeand the final prediction function, construct a new combined decision tree algorithm AdaTree.WL.By contrast with the SVM algorithm, found the advantages and disadvantages of the twoclassification algorithms. Studies show that the improved decision tree classification algorithm isbetter than SVM algorithm in the overall classification accuracy, but for a single land type, bothhave advantages and disadvantages: AdaTree.WL algorithms is better in extracting mostvegetation types, and SVM is better in the extraction of artificial types.(2) For high-resolution image always has a large number of features, there are exitingredundant problems. This paper constructs two feature selection models to solve it. Firstly, use adecision tree classifier CART (Classification and Regression Trees) to calculate the contributionof features in classification, and through a large number of experiments to screening a set offeatures initially. Then, gradually eliminate features with large correlation and interference usingfeature selection model based on the correlation characteristics, and apply the feature set toverify the classification results with examples. Experiments show that the optimized feature setcan greatly improve the classification accuracy of vegetation types.(3) This paper studies the relationship between the the unique red edge band and vegetationclassification. Experiments show that when the red edge band is added in, the accuracy of GLCand SVM classification was increased by2.45%and10.12%. In addition, the paper also presentstwo vegetation classification method based on multi-temporal features and texture features. Approach based on multi-temporal approach allows GLC and SVM classification accuracy wasincreased by7.47%and6.29%. Approach combined texture classification method makes GLCand SVM classification accuracy was increased by3.91%and2.56%.