Detection Of Abnromal Training Samples In Remote Sensing Images Based On Median Absolute Deviation And Self-training Convolutional Neural Networks

Remote sensing image classification has always been one of the important research directions in the field of geomatics and remote sensing,and its classification results provide basic resources for many natural environment and socio-economic applications.In the supervised classification of remote sensing images,the quality of training samples is directly related to the results of remote sensing image classification.However,whether in pixel-level remote sensing image classification or scene-level remote sensing image classification,there may be abnormal training samples in the selected training samples,which will have a certain impact on subsequent research and applications.If an appropriate method is adopted to detect and eliminate the existing abnormal training samples,the classification accuracy of remote sensing image will be improved.Therefore,it is a great significance to choose an appropriate anomaly detection method.In outlier detection,the performance of different detection methods is different,the detection results of different data are also different and there is usually a lack of systematic comparison and analysis.In pixel-level remote sensing image abnormal training sample detection,many methods are only effective when the training data is polluted by a small number of abnormal samples.In scene-level remote sensing image abnormal training sample detection,the characteristics of scene images are relatively complex.If the sample difference is converted into point distance or sample reconstruction,the anomaly detection effect is not ideal.The convolutional neural network method that can effectively detect requires a large number of training samples as support.In response to the problems in the above research,this master’s dissertation will carry out research from the following three aspects:(1)Three common outlier detection methods are introduced,namely Z-score method,boxplot method and median absolute deviation(MAD)method.Then three methods be used to detect outliers in simulated data sets and real-life data sets.In the first set of simulation experiments,the magnitude of outliers is fixed but the number of outliers is varied systematically.In the second set of simulation experiments,the number of outliers is fixed but the magnitude of outliers is varied systematically.In the real-life experiment,two sets of real-life observation data sets are selected for the experiment.The results show that the outlier detection effect of MAD method is better than Z-score method and boxplot method in both simulation and real-life experiments.(2)The MAD method is used to detect abnormal training samples in pixel-level remote sensing image classification.The abnormal training samples are divided into two types,i.e.,impure samples and wrongly selected samples.According to the principles of variance inflation and mean shift,the sum of standard deviation(for impure samples)and the sum of mean(for wrongly selected samples)for two type of training samples are calculated separately.Then the decision values can be obtained by using the MAD method and the abnormal training samples can be detected and eliminate.The training samples before and after removing abnormal are used for SVM classification,and the classification results of overall accuracy and Kappa coefficient are compared.The results show that the MAD method can accurately detect and eliminate impure and wrongly selected training samples.The classification accuracy by using MAD method can be improved effectively.(3)In remote sensing scene images,the self-training algorithm can use a small amount of labeled data to train a large amount of unlabeled data,thereby achieving the purpose of expanding the training dataset.This method can make up for the insufficient performance of convolutional neural networks when there are few training samples.A combination of self-training algorithm and convolutional neural network is proposed to detect abnormal training samples in two set of remote sensing scene datasets,i.e.,SIRI and RSSCN,which contain abnormal training samples after processing.The results show that the accuracy,precision,recall and F1-score of the three self-training convolutional neural networks(i.e.,self-training Google Net,self-training Res Net and self-training Dense Net)in the two datasets can reach more than 93.8%,indicating that the effect of training convolutional neural network in scene-level remote sensing image anomaly detection is better.