Multiple Instance Learning Methods For Representation Of Target In Hyperspectral Remote Sensing Images

Recently,hyperspectral remote sensing images have played a more and more important role in practical tasks such as military reconnaissance,mapping,environmental monitoring,geographic information systems,and precision agriculture.In these realworld applications,target detection and classification becomes the key procedure.Then,the accurate representation of hyperspectral target,which is the main problem in these tasks,tends to be the key issue.Therefore,almost all target detectors require a well-characterized target representation spectrum,but the common hyperspectral target representation methods have limited representation capabilities.In recent years,many researchers try to use machine learning-based methods to learn more powerful feature representation from the targets.Among these machine learning methods,multi-instance learning methods have become the effective ones for studying hyperspectral target representations since they do not require precise pixel-level semantic labels.However,in the multi-instance learning method for hyperspectral target representation,due to the limited number of positive samples,there exist two problems which lead to worse representation of the learned features of the target: one is that the target example in the positive package is usually far less than the background example,which leads to the unbalance problem of example-level data.The other is that the number of the extracted positive packets is much smaller than that of the negative packets,which generates the bag-level data imbalance problem.The above-mentioned problems show negative effects on the representation of the targets.In addition,due to the large amount of data and the redundancy of hyperspectral image data,there exists the “dimensionality disaster problem” in the target representation process.Therefore,this paper focuses on the multi-instance hyperspectral target representation methods to solve the example-level and packet-level data imbalance problems in multi-instance learning,introduces the data dimensionality reduction method to solve the problem of large data redundancy to improve the accuracy of target expression,and finally validates the effectiveness of these methods with realworld hyperspectral target detection problem.Firstly,an example-level equilibrium data multi-instance learning method is proposed for the example-level data imbalance problem.The positive example set,which is most likely positive in each package,is extracted.Based on the positive example set,a new positive sample is synthesized,and the positive sample is added.The proportion of samples in the positive package improves the ability to express hyperspectral targets.The validity of the proposed method is verified on real-world hyperspectral data.The experimental results show that the proposed method makes the composition of the positive packet more balance,so as to learn a more accurate target representation and improve the performance of target detection.Secondly,a multi-instance learning method for packet-level equalization data is developed for the packet-level data imbalance problem.The samples with the most likely positive and most likely negative in each packet are combined to form a corresponding positive and negative example set,and new positive and negative samples are synthesized.Add to the empty package with the positive label prepared in advance,that is,synthesize the new package.By increasing the proportion of positive packets in the test set,the data samples are balanced to improve the ability to express hyperspectral targets.The effectiveness of the proposed method is verified on real-world hyperspectral data.The results show that the proposed method makes the number of positive and negative packets in the test set more balance,so that the more optimized target representation is learned,and then the performance of target detection is also improved.Finally,a multi-instance learning method,which is based on PCA dimensionality reduction,is proposed to improve the "dimension disaster" problem caused by hyperspectral data redundancy.By performing the example-level PCA dimensionality reduction for the extracted positive and negative packets,the high-dimensional data is passed.The cross transformation transforms to the appropriate low dimension,removing unnecessary information and leaving only valuable dimensions.Therefore,the representational ability of the hyperspectral targets is improved.The effectiveness of the proposed method is verified on real hyperspectral data.The results show that the proposed method also improves the performance of target detection.