Research And Application Of Spectral-Spatial Hyperspectral Image Classification

With the rapid development of society,people’s demand for macro perception and deep cognition of space is also increasing day by day.Hyperspectral Image(HSI)contains rich spectral and spatial information,which can provide sufficient data support for fine ground object recognition.HSI is increasingly being used in a variety of commercial,industrial and military applications,such as precision agriculture,mineral exploration,etc.HSI classification algorithms are the technological cornerstone of these applications.Taking precision agriculture as an example,HSI-based crop fine classification is an important basis for crop growth monitoring and yield prediction,which is of great significance for planting structure supervision and food security guarantee.However,many of the traditional HSI classification methods require sufficient training samples to achieve desired classification performance,and are faced with the problem of large consumption of computing resources.How to strike a balance among limited training samples,high classification performance and low computing resource consumption is one of the most important research directions in the field of HSI classification.In view of the above-mentioned problems,two spectral-spatial HSI classification methods are proposed,and integrated application for crop fine classification is carried out.The main research work can be summarized as follows:(1)Regarding the small sample size problem,a spectral-spatial HSI classification method based on Random Multigraphs(RMG)is proposed.RMG is a graph-based ensemble learning method that creates graphs by randomly selecting feature subsets,which is rarely used in existing HSI classification methods.RMG can deal well with the problem of small sample size classification problem,which is very common in HSI applications.Firstly,the original HSI is dimensionally reduced by Maximum Noise Fraction transformation and spectral features are extracted.Then,a series of spatial features were extracted based on Local Binary Pattern algorithm.Then,these spatial features and spectral features are stacked to obtain multiple space spectral feature vectors,and these spectral-spatial feature vectors were input to RMG for classification.The optimal parameters and classification results can be obtained by voting.With only10 training samples for each type of ground feature,lateral comparison experiments of association methods were carried out on three real hyperspectral datasets: Indian Pines,University of Pavia and Salinas Scene.Experimental results show that the overall classification accuracy of the proposed method is improved by 28.6%,36.78% and68.66%,respectively.(2)In view of the problem that adjacent features are easily confused in the HSI classification process,a spectral-spatial HSI classification method based on Sparse Representation was proposed.Firstly,to enhance the control of the boundary information of adjacent ground objects,a weighted Laplacian smoothing constraint vector is constructed,which is composed of the target pixel and its four nearest pixels by weighting factors.Secondly,a block sparse dictionary is constructed based on the above vector,and multi-scale spatial information is effectively utilized through an adaptive strategy.Finally,a sparse constrained optimization problem can be solved with Simultaneous Orthogonal Matching Pursuit algorithm to achieve ground object classification.The optimal parameters and classification results can be obtained by voting.Compared with some existing representation-based HSI classifier,the proposed method achieves better HSI classification accuracy on three real HSI datasets,which verified the effectiveness and superiority of the proposed method.(3)Based on the theoretical research of(1)and(2),a HSI classification technical system suitable for crop fine classification is constructed.The experimental duration and memory consumption of the proposed method are compared with those proposed in(2),and the good performance of the new technical system in fine crop classification is verified.Based on the airborne HSI collected in the typical agricultural area of Jianghan plain,different application scenarios are simulated from three aspects of different levels of Gaussian white noise,different levels of colored noise and different number of training sets,and the applied research to crop fine classification is done.On the whole,the classification is relatively ideal,and the most misclassified typical regions are mainly concentrated in the areas with relatively dense crops.In the process of subsequent application and promotion,the following two aspects can be adopted to avoid the occurrence of the above situation as much as possible:(a)a relatively larger number of training samples should be selected for the areas with smaller planting plots and dense crops;(b)in the process of data acquisition,sensors with higher spectral and spatial resolution can be adopted.