The Research On Ensemble Method And Its Application On Remote Sensing Image Classification

Ensemble methods are characterized by good generalization and high stability. It is widely used in complicated classification task. In this paper, we do some research on ensemble method and its application on remote sensing image classification. Rotation Forest ensemble method would over-fitting when classify the high dimension data. In order to overcome the problem, an improved Rotation Forest method which is called RF-ELM is proposed. RF-ELM use Rotation Forest to increase the diversity among the base classifiers. Then we choose Extreme Learning Machine as base classifier. Because the fast training process and good adaptability of ELM, RF-ELM could get good classification results. As remote sensing images are nonnegative, a novel remote sensing image classification method based on neural network ensemble is proposed and carried out by two steps:select nonnegative matrix factorization to extract features and use Q statistic to measure the diversity in ensemble method. In addition, traditional remote sensing image classification methods use spectral characteristics only while ignoring the other features, which would meet metameric substance of same spectrum and metameric spectrum of same substance. To solve the problem, we proposed a remote sensing image classification method based on multiple feature combination. Gabor transform is used for extracting texture feature of the image. In order to obtain classifiers with large diversity and promote the classification accuracy, we apply an ensemble algorithm to integrate the spectrum feature and texture feature after feature segmentation. The simulation experiments on the UCI data and real remote sensing images show that the proposed method can achieve good classification results.