Spectral-Spatial Methods And Collaborative Learning For Hyperspectral Image Classification

Hyperspectral imagery(HSI)can be regarded as the sampling data of the target scene object acquired by the imaging spectrometer with the corresponding spatial and spectral resolutions,which realizes "an image-spectrum merging technology" of the data.HSI imaging technology has been widely used in urban planning,water resources and wetland protection,precision agriculture,geological hazard forecasting,forest vegetation survey,military target detection and identification.HIS classification is one of the main tasks of HSI data processing,and refers to labeling each pixel in HSI scene data given predefined feature categories and learning models,and identifying each pixel as the corresponding category according to the characteristics of the spectral features.HSI classification is the premise for accurate interpretation and quantitative analysis of subsequent hyperspectral data.It is of great practical significance to study the classification models and algorithms and find more efficient algorithms for HSI classification and recognition.The dissertation starts from the two aspects,spectral-spatial classification of hyperspectral remote sensing imagery and the scenario of limited available training samples,and solves the problem of over-smoothness/under-correction,label bias balance and detailed edge information protection in Markov random fields framework and collaborative learning scheme,and furthermore improves the classification performance HSI land-cover objects and protects more edge details.This dissertation has the following characteristics and technical contributions:(1)This dissertation proposes a novel Markov random fields(MRFs)method integrating adaptive inter-class-pair penalty(akCP2)and spectral similarity information(SSI)for HSI classification.aICP2 structurally combines K(K-1)/2(K is the number of classes)classical'Potts model's with K(K-1)/2 interaction coefficients.aICP2 tries a new way to solve the key problems,insufficient correction within homogeneous regions and over-smoothness at class boundaries,in MRF-based HSI classification.It is assumed that different class pairs should be assigned with various degrees of penalties in MRF smoothness process,according to pairwise class separability and spatial class confusion in raw classification map.Fisher ratio is modified to measure pairwise class separability with a training set.And,gray level co-occurrence matrix(GLCM)is used to measure spatial class confusion degree.Then,alCP2 is constructed by combining Fisher ratio and GCLM.aICP2 applies larger penalty on class pairs that confuse with each other seriously to provide sufficient smoothness,and vice versa.In addition,to protect class edges and details,SSI is introduced to make the penalty of related neighboring pixels small.aICP2ep denotes the integration of aICP2 and SSI-based edge preserving(EP).The further improved method is both inter-class-pair and inter-pixel adaptive.A graph-cut-based ??-swap method is introduced to optimize the proposed energy function.Experimental results on real HSI data indicate that the proposed method outperforms compared MRF-based and other spectral-spatial approaches in terms of classification accuracies and region uniformity.(2)This dissertation proposes a novel MRFs-based scheme considering label bias(LB)balance,label smoothness,and edge preserving(EP)for HSI spectral-spatial classification.MRFs are widely used to incorporate label smoothness prior in HSI classification.However,few studies noticed the pattern of un-smoothness phenomena in raw land-cover mapping,and its effect on the refined result.We start with a problem that within some locals,where most pixels are mislabeled,classical Potts model does not consider whether the shared label of adjacent pixels after regularization is correct or not,and may change nothing,or even cause worst results.This paper assumes this due to LB in raw mapping results.For this,a novel prior,LB estimate(LBE)is proposed and defined to measure LB.Then,a novel term,LB balance(LBB)with adaptive intra-class penalty(aICP)is derived and integrated in MRF model to balance and offset LB.Moreover,an EP strategy is utilized.LBBep indicates the proposed scheme combining LBB,label smoothness,and EP.A graph-cuts-based a-expansion algorithm is introduced to solve the proposed energy function.Experiments on two groups of real HSI data indicate that LBBep outperforms other spectral-spatial and MRF approaches in region consistency and accuracy metrics.(3)This dissertation proposes an adaptive edge preserving(aEP)scheme in MRFs for spectral-spatial classification of HSI.Two binate problems have limited the performance after MRF regularization,over-smoothness at class edges and insufficient refinement within homogeneous regions.This work divides and conquers the problem class-by-class,and integrates K(K-1)/2(K is the class number)aEP maps(aEPMs)in MRF model.Edge detectability measure(EDM)is proposed and estimated via the training set.And,for each pair of classes,aEPM is optimized by maximizing interclass EDM and,meanwhile,minimizing intraclass EDM.The proposed aEPMs is integrated with Potts model to regularize the pixelwise labeling by pure spectral and spectral-spatial methods,respectively.The graph-cuts-based ??-swap method is modified to optimize the designed energy function.Additionally,to further evaluate the final results at class borders,segmentation evaluation metrics are introduced.Experiments on real HSI data denote the superiority of aEPMs in evaluate metrics and region unity,especially in detail preserving.(4)This dissertation proposes a novel collaborative learning(CL)framework for HSI classification,in which active learning(AL)and semi-supervised learning(SSL)are collaboratively integrated using clustering(CLUC).Recently,collaborative learning(CL)is introduced to combine AL with SSL,and solve the problem of limited training samples.CLUC attempts to obtain more diversity and higher confidence of additional training samples in both AL and SSL.Note that clustering methods,which are used separately to enhance AL or SSL,are utilized to integrate these two learning processes in CLUC.First,all unlabeled samples are assigned into clusters.Based on the clustering result,clustering-based query(CBQ)for both AL and SSL,and CBQ-based pseudo-labeling(CBQPL)for SSL are designed for CLUC.Second,the most and secondary uncertain samples in each cluster are selected by CBQ for AL and SSL,respectively,to ensure their informativeness.Third,CBQPL assigns the selected secondary uncertain samples with the same label as the most uncertain one,which is manually-labeled in AL within the same cluster.CBQPL makes the confidence of pseudo-labeling rely on the clustering results.We evaluate the performance of CLUC on three real HSI images.The performance of the proposed method is tested under different numbers of labeled samples and compared with several approaches.We can observe from the experimental results that CLUC have superiority in classification maps and objective metrics with limited training samples.