Study On Classification Of Remote Sensing Image Based On Improved Extremely Randomized Clusting Forests

The classification of remote sensing image is the most fundamental and important task for remote sensing.The actual needs of Environment, resources, socioeconomic and military put forward higher requirements on the theory, technology and methods of Classification. Massive advancement had been achieved in theory and technology during past decades; however, the classification can still not full met the demand of practical application. Because many factors, such as the complexity of the landscape in study area, selected remotely sensed data, and image-processing and classification approaches, may affect the success of a classification, classifying remotely sensed data into a thematic map remains a challenge.Aiming toward improving the accuracy and efficiency of classification and meting the need of practical application, the extremely randomized clusting forests (ERC-Forests) algorithm in machine learning was introduced and improved in many aspects. A systemic and synthetically research was concerned on extremely randomized clusting forests algorithm and the application in multi-spectral image, hyperspectral image and object oriented image classification.The main works and innovative points are show following:(1) In the paper, the extremely randomized clusting forests were imported, and many improvements were made. For example, in order to avoid the appearance of unbalance decision trees, the tree balance index were added in the algorithm. For the purpose of increasing the efficiency of algorithm, the impurity of leaf node was amended from0to0.05. For fear of unnecessary accuracy loss, the decision tree level was set to fifteen. To reduce the computational complexity of algorithm, combine with the attribute of ERC-Forests, the prepruning technology was handled.(2) The experiment with multi-spectral image was conducted to compare the accuracy and the efficiency of original and improved extremely randomized clusting forests. The experiment indicates that the accuracy was increased by1%,and the efficiency was increased by23.4%(3) The improved ERC-Forests algorithm was firstly used in the multi-spectral image classification, and compare with five classical classification approaches in experiment, and choose the feature extraction index. The results of experiment show that the accuracy of improved ERC-Forests algorithm was increased by4%than maxlike likelihood,26%than Parallelepiped,16%than M distance and14%than minimum distance.(4)After the study on the advantage and the shortage of existing hyperspectral image classification method, this paper put forward a new approaches for hyperspectral image classification. Comparing with widely used approaches, such as maximum likelihood (ML), support vector machine (SVM) and neural network (NN) and so on, we can draw a conclusion that in comparison with the other methods, the classifier base on the improved ERC-Forests algorithm has the following advantages, such as simple structure, easy training, the high accuracy etc..(5)The relate theory of object oriented methods was studied, and research on the combination with the improved ERC-Forests algorithm and eCognition was discussed. After the image segmentation with eCognition, the classification was conducted on the exported results of eCognition with the improved ERC-Forests algorithm; the results indicated that the improved ERC-Forests algorithm can effectively ameliorate the results of classification by5%in accuracy.In conclusion, when the ERC-Forests algorithm were used in multi-spectral image classification, hyperspectral image classification and object-oriented image classification, the accuracies of these classifications are ideal. In other words, the ERC-Forests algorithm is suitable for optical and near infrared remote sensing image classification.