Remote Sensing Image Detailed Classification Based On Improved Adaptive Boost Algorithm

The classification of remote sensing images has always been a hot research topic of remote sensing information extraction.Due to the complexity and limited technology of ground objects,the classification of ground objects with similar spectral characteristics and spatial shape characteristics is a difficult point in our research.Based on this,an integrated learning algorithm(AdaBoost)is introduced in this paper to solve the problem of fine classification of medium-resolution remote sensing images.Due to the low separability and high noise content of the middle resolution remote sensing image data,and the fact that the AdaBoost algorithm is sensitive to noise and prone to overfitting,an improved adaptive lifting algorithm model(NP-AdaBoost)is proposed in this paper,and the results prove that the improved algorithm model is suitable for fine classification in the study area.The main research methods and results are as follows:(1)Improving adaptive boosting algorithm.For noisy samples,considering the concept of a full probability formula for sample noise,this paper divides the sample data after training into four categories.In one experiment,Correct sample classification and wrong sample classification form a complete event group.Under the condition of correct classification and wrong classification,the probability that the sample is noisy is equivalent to a conditional probability,therefore.this paper constructs a full probability model of prediction data noise and the noise model is used to predict the likelihood of the training sample data for the noise.The results of the model as a new parameter of the adaptive boosting algorithm.It is proved theoretically that when the sample weight is updated and the classification is correct,the sample data with large noise will decrease more than that with small noise.In the case of classification error,the sample data with large noise will increase less than that with small noise.(2)Verifying the effectiveness of the improved algorithm.The experimental data are simulated and the data set is from the UCI database for machine learning.Due to similar terrain features,low separability,middle resolution remote sensing images affected by noise environment,so data sets with different attribute characteristics are constructed to simulate different separable environment,data sets with different noise proportions are constructed to simulate different noise environment.then this paper compares the classification effect of the original algorithm with that of the improved algorithm.Through accuracy analysis,training error analysis and test error analysis,the results show that the NP-AdaBoost algorithm can make the trend of training error more stable and reduce the overfitting phenomenon in the experiment,so as to reduce the test error.(3)Classfying objects in research area.Selecting 2013 Landsat8 OLI remote sensing image,six kinds of vegetation index and the red and near-infrared single band characteristic,The paper classifies woodland and grassland.Finally comparing with other supervised classification methods(maximum likelihood,minimum distance,mahalanobis distance and decision tree),the improved adaptive boosting algorithm is a kind of can be fine for remote sensing image classification and effective algorithm with higher accuracy.Compared with the original algorithm,the Kappa coefficient is improved by 5.5%,and the overall accuracy is improved by 2.8%.