Cross-Scene Feature Selection For Hyperspectral Images

Hyperspectral images(HSIs)are captured with high spectral resolution,and rich spectral information of HSIs can be utilized for classification of land-cover objects.However,HSIs contain noisy bands and information redundancy,which lead to difficulties for subsequent applications.Therefore,it is necessary to reduce redundant dimensionality.Feature(band)selection is a common method for dimensionality reduction in HSIs data,it selects a subset from the original feature set without changing the data structure,thus retaining the original physical meaning of the features.Traditional feature selection algorithms can only be performed on a single hyperspectral image(scene),while cross-scene feature selection has not been extensively studied.This dissertation studies cross-scene feature selection for hyperspectral remote sensing images.This problem is carried out on different scenes(datasets),and there is a problem of spectral drift(data difference),the traditional single scene based feature selection algorithms is no longer applicable.In order to solve this problem,this dissertation proposes two cross-scene feature selection algorithms.These two algorithms make full use of the information of the source and target scenes,considering two important factors: a)the discriminant of the selected feature subset on the land-cover objects,b)the consistency(invariance)of the selected features between different scenes.The specific research contents of this dissertation are as follows:(1)Feature selection based on cross-scene information gain algorithm.The algorithm considers two kinds of information gains for the feature,one is class-specified information gain,and the other is domain-specified information gain.This dissertation proposes a cross-scene information gain algorithm based on above two kinds of information gains.The selected features by the algorithm have discriminant to separate different land-cover classes,and also deliver the consistency between different scenes,so as to reduce the influence of spectral drift.At the same time,for the practical application on hyperspectral images,this study introduced a calculation method of information gain on continuous numerical features.(2)Feature selection based on cross-domain Relief F.The algorithm makes full use of the information of the source and target scenes to ensure the consistency of selected features between the scenes.In details,after randomly selecting a sample from the target scene,the nearest neighbor samples are selected from the source scene and the target scene respectively,and the weights of the features are updated by cross-scene rules.Thereby,a feature subset is selected that contributes a lot to the separation of classes as well as consistency between the scenes.In this dissertation,the above two algorithms are evaluated in the cross-scene HSIs.The experimental results show that the proposed cross-scene feature selection algorithm can effectively improve the classification accuracy of the target scene.