Hyperspectral Image Classification Based On Spatial Spectral Features And Decision Fusion

Hyperspectral image（HSI）has the characteristics of wide coverage,many bands and a large dataset.Therefore,it is possible to accurately classify the ground objects by using HSI in modern agriculture,resource protection,ocean monitoring,urban remote sensing and other fields.However,there are many challenges and difficulties in feature extraction and classification applications,because of the large amount of data and strong band correlation.This paper conducts research from three aspects of hyperspectral image dimensionality reduction,spatial-spectral feature extraction and multi-classifier decision fusion,in which the improved methods and strategies are proposed respectively.The efficiency and feasibility of the proposed methods are confirmed by experiments.The following are the three main research contents:（1）An adaptive collaborative graph discriminant analysis（ACGDA）method for hyperspectral image dimensionality reduction is proposed.In order to maintain the inherent geometric structure of the original data and improve the interpretability of the underlying image,the intrinsic relationship between the pixels in the modeling class is defined in the process of graph construction.ACGDA combines the distance-weighted Tikhonov regularization with the collaborative representation based onl₂-norm minimization,thus generating stronger recognition ability.In addition,the graph weight matrix is designed in the form of a block-diagonal structure,reducing the computational cost and further improving discriminative power.（2）A new image classification method based on local binary pattern（LBP）is proposed.This method segments all bands to several band groups firstly and employs principal component analysis（PCA）to extract the main spectral features of each group.Then,local binary pattern is used to extract spatial texture features and support vector machine（SVM）classifier is used for classification.Finally,the sub-classification results are fused by majority vote（MV）strategy.Experiments on two typical hyperspectral datasets show that the proposed system improve the classification accuracy significantly and has good robustness for small-sample-size situation and noise environments.（3）Aiming at the disadvantages of traditional muti-classifier decision fusion system,two adaptive weight decision fusion strategies（adjustMV and adjustLOGP）are studied and an adaptive decision fusion algorithm for hyperspectral image classification is proposed.Traditional decision fusion strategies use unified weight coefficients for decision fusion,but the adaptive weight decision fusion strategies proposed in this paper reduce the influence of poor performance of sub-classifiers and noise bands by assigning bigger weights to the sub-classifiers with better performance.For adaptive decision fusion classification algorithm,the hyperspectral images bands are grouped according to the correlation coefficient matrix,and then spatial-spectral joint-feature extraction is performed on each group of bands.Next,gaussian mixture model（GMM）or support vector machine（SVM）classifier is employed to classify each group of spatial-spectral features.Finally,two improved decision fusion strategies are used to perform decision fusion on the classification results of the sub-classifiers,so that the band groups with low classification accuracies and outliers have the least impact on the final classification results.