Research On Hyperspectral Image Classification Methods With Noisy Labels

Hyperspectral image(HSI)classification is one of the most classical research topics in the field of remote sensing images.Improving both classification performance and computing efficiency are of great significance for many fields,such as precision agriculture,military anticounterfeiting,and urban planning.Therefore,it is still one of the most popular research topics for many researchers to design an HSI classification method with high accuracy,high timeliness,and high robustness as well.However,the existing HSI classification methods are designed based on the assumption that the training samples are labeled exactly.Due to the existence of other objective factors,it is extremely hard to meet the above assumption in the actual labeling process,which causes some samples to be labeled incorrectly in the ground truth(i.e.,noisy labels).In view of the above problem,this article considers the noisy labels as abnormal data in the entire training sets.It is intended to reduce the impact of HSI classification methods from the noisy labels by designing a variety of key noisy labels detection and elimination methods.Among them,the main works of this article are summarized as follows:The collaborative representation(CR)anomaly detection algorithm mainly estimates and compares the residual information of each pixel to determine the degree of abnormality of each pixel.Because of its excellent detection performance,CR algorithm is widely used in the abnormal detection of HSI.Therefore,based on the introduction of the residual metric-based CR algorithm,this paper proposes a class-dependent collaborative representation(CDCR)HSI classification method with noisy labels for detecting and identifying the noisy labels from the HSI.The specific steps are shown as follows: First,we need to calculate the Tikhonov regularization-based weight matrix for each class of training samples,and roughly estimate the degree of difference among training samples.Then,the CR algorithm is used to obtain the residual information of each class of training samples.Experimental results show that the larger the samples with residual value,the greater the probability of being identified as noisy labels.Finally,this characteristic is used to effectively detect and eliminate noisy labels through a designed decision function.Experimental results show that the method can effectively detect noisy labels in the HSI and improve the performance of HSI classification methods.Density peak(DP)clustering algorithm can accurately obtain the density information of each pixel in the HSI,and then effectively measure its degree of dispersion and anomaly.Therefore,this paper proposes a method based on DP clusterin algorithm for the HSI classification with noisy labels.First,the correlation coefficient(CC)measurement algorithm is used to calculate the spectral distance among training samples.Then,the local density information of each training sample is obtained using the DP algorithm.In general,noisy labels tend to have lower density information.Finally,a decision function based on local density is used to detect noisy labels.The experimental results performed on multiple real hyperspectral data sets show that the method can effectively use the density information of the sample to accurately measure the degree of dispersion and anomaly of the sample.Moreover,it can not only achieve more excellent and robust noisy labels detection performance based on high detection efficiency,but also further improve the classification performance of the HSI.Although DP-based HSI classification method with noisy labels can obtain high detection results,it does not consider the spatial context information of each pixel.Therefore,to further improve the detection accuracy of noisy labels,this paper designs an algorithm for the HSI classification with noisy labels based on spatial density peak(SDP)clustering.Specifically,the proposed method consists of the following steps: First,we need to obtain the local neighbor samples around each training sample in the square windows.Then,to make full use of the sample spatial context information,some samples in the local neighborhood are selected to measure the central sample as candidate samples,and then we can calculate the CCs between each class of training samples.Here,we introduce two different selection rules of neighbor samples in the local neighborhood as SDP and K-SDP.Then,the local density information of each training sample can be calculated,with using the obtained CC matrixes as the prior knowledge of the DP algorithm.Finally,the decision function based on local density is used to effectively detect noisy labels with low local density values.Experimental results on four widely used hyperspectral data sets demonstrate that it is a more effective approach to detect noisy labels by introducing spatial information.In summary,in order to detect the noisy labels in the HSI,this paper designs a variety of detection methods by using different kinds of measurement algorithms.The experimental analysis proves the detection performances and robustness of the proposed algorithms.Moreover,this paper provides some novel solutions for the HSI classification with noisy labels.