Research On High Accuracy Classification For Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing images can reflect the physical properties of objects and exploit sophisticated and abundant spectrums to observe objects.It is precisely because of high spectral resolution that hyperspectral remote sensing images have been widely used in military reconnaissance,mineral exploration,agriculture precision mapping,environmental monitoring and other fields at home and abroad.It is the premise of accurate analysis and interpretation that deeply studies the characteristics of hyperspectral data aiming at useful feature enhancement and unwanted components elimination.In addition,hyperspectral remote sensing images classification is the foundation of recognition and detection.However,it is confronted with many challenges for achieving high precision hyperspectral images classification:how to model the deep characteristics of hyperspectral images to solve poor classification performance with limited training samples;how to solve the common problem of the identical material with different spectral signatures for increasing classification accuracy;how to increase the difference of interclass and decrease the difference of intraclass,etc.To solve these problems,this dissertation deeply analyzes the characteristics of hyperspectral images and systematically reviews the existing literatures.Multiple novel feature enhancement and classification technologies for hyperspectral images are proposed in this dissertation.The proposed methods are evaluated by three real hyperspectral remote sensing images.The main achievements of this dissertation can be summarized as follows:(1)The commonly used classification methods based on local optimization cannot model the deeply spatial and spectral structures in hyperspectral images,and thus,not obtain an effective classification performance when the quite limited training samples of object.In order to deal with this problem,this dissertation deeply analyzes the effect of convolutional operation on hyperspectral images and pioneers proposing a deep feature enhancement method based on convolutional neural network(CNN)for removing noise and enhancing the salient spatial structures,shapes and boundaries.In fact,this method can achieve decreasing similarities of the pixels belonging to different objects while increasing similarities of the pixels belonging to the identical object.The results of feature extraction method are input into different classifiers including support vector machine(SVM),extreme learning machine(ELM)and three sparsity-based classifiers.Experimental results demonstrate that the features obtained by the proposed method can well represent the spatial contextual information of objects,and thus,can achieve breakthrough performance when identifying the object with limited training samples.For example,the object with only 20 pixels in Indian Pines image can achieve 100%classification accuracy.(2)As the key characteristics of HSIs,spectral reflectance is usually affected by sensing mechanism,noise and capturing circumstance,and thus,resulting in the identical material with different spectral signatures.It is deeply analyzed that the seriously distorted pixels are usually located in the boundaries among objects because these pixels can be more easily affected by the surrounding objects.Thus,this paper introduces alpha channel into hyperspectral remote sensing images and proposes a novel feature enhancement method based on the closed form matting model.The defined alpha channel is served as guidance image and the hyperspectral remote sensing image is decomposed into hyperspectral foreground and hyperspectral background with the assumption of local smoothness in foreground and background.Finally,the alpha channel,hyperspectral foreground and background are linearly combined to obtain a feature enhanced hyperspectral data,which effectively overcome the problem of the identical object with dissimilarities spectral characteristics.Experimental results indicate that the extracted features have positive influence on various classifiers.(3)Owning to hyperspectral remote sensing images with high dimensions,it is difficult to realize real time processing of hyperspectral feature enhancement.It is analyzed that the seriously distorted pixels are usually located in the boundaries so that this paper introduces guided filtering that can well preserve boundary details.Especially,guided image filtering can achieve real time processing.The guidance image is defined by statistical and morphological features,firstly.The computational time is dramatically decreased by down sampling the linear coefficients.In the reconstruction procedure,the guidance image with full resolution can faithfully guide the hyperspectral images.Experimental results demonstrate that the fast feature enhancement algorithm based on guided filtering can solve the phenomenon of same material with different spectral signatures and the high dimensional data with limited training samples can achieve 100%classification accuracies.The classification accuracies of other classes also have significant improvement.Especially,the processing time of hyperspectral image enhancement is dramatically decreased.It is shown that the guided filter based fast feature enhancement algorithm can be simply sped up from O(N)to O(N/s~2)time with a down sampling ratio s.