A Study Of Label Noise Based On Ensemble Learning In The Classification Of PolSAR Images

Polarimetric Synthetic Aperture Radar(Polarimetric SAR for short) is muti-parameter,multi-channel imaging Radar system. It is widely used becaused of its advantage of all day,all weather and high resolution. Polarimetric SAR image classification methods based on machine learning has achieved a high classification accuracy.But the classification results will be affected of the presence of label noise. For the problem of label noise in polarimetric SAR image classification,reseach is carried,mainly including the following three aspects:1. Combined with polarimetric SAR scattering features and texture features,proposed polarimetric SAR image classification algorithm based on Ada Boost(K nn.Ada Boost). The new method uses polarimetric SAR scattering features and texture features as input features of Ada Boost. K nn.Ada Boost algorithm use KNN to compute the anti-noise factor of each pixel of polarization SAR image, then present a new sample weights update strategy by K nn calculations. Experiments using multiple sets of polarimetric SAR data.Experimental results show that, Knn.Ada Boost algorithm improves the classification accuracy and has good anti- noise performance.2. Based on K nn.Ada Boost,a new polarimetric SAR image semi-supervised classification algorithm,named Semi.Knn.Ada Boost, is proposed. In the framework of Knn.Ada Boost, use wishart distance, in the end of each iteration, using labeled samples to obtain wishart cluster centers, select the nearest m samples and add them to their corresponding cluster for the next iteration. Experiments using multiple sets of polarimetric SAR data, the results show that, Semi.Knn.Ada Boost enrich the training samples, classification accuracy has some improvement.3. A method for estimating the noisy level is proposed based on ensemble learning in the polarimetric SAR image classification problems, EEL(estimated by ensemble learning). Taking coherent matrix of Polarimetric SAR images as the feature vector,then different classification methods be used to get independent classifiers. And then judge the label of a sample is noise or not,by using the strategy of majority vote.Ifmore than m classifier classification results is the same,we identify the label is noisy label,otherwise,it is not.Experiments with UCI data and a sets of simulated polarization SAR data,it is showed that the proposed method work well when the noisy level is low.But when the noisy level is high,the result is not satisfied.Our work is supported by national natural science foundation of C hina(No.61173092), new century excellent talents to support plan(No. 66ZY110) and shaanxi province science and technology research and development projects(No.2013KJXX-64).