Research On Classification Method Of Hyperspectral Remote Sensing Images

In recent years,in the field of remote sensing image research,hyperspectral remote sensing image has attracted great attention because of its unique advantages.It can not only have high spectral resolution,but also have high spatial resolution.Its "map-inone" nature can better reflect the natural attributes of the image,and can provide rich information for the recognition of ground objects.Therefore,in order to make better using of hyperspectral remote sensing data to obtain useful geographic information,people begin to pay attention to the classification of hyperspectral data.Classification is a very important basic work.High-precision remote sensing image classification is of great significance to meet the needs of social and economic development.Although there have been many hyperspectral remote sensing image classification methods,due to the high spectral resolution of hyperspectral data itself,it will inevitably lead to high correlation between bands,and its non-linear and highdimensional characteristics make the existing ones.Many classification methods are not suitable for hyperspectral data,which results in the inaccuracy and inefficiency of many existing classification methods for hyperspectral remote sensing images,and it is difficult to meet the needs of practical applications.In view of this,this paper focuses on the research and implementation of three common classification methods for hyperspectral remote sensing image classification,including the traditional statistical pattern-based classification method,the classical Support Vector Machines(SVM)classification method and the sparse representation Classification(SRC)classification method which is currently a hot research topic.Through the classification result maps and classification evaluation indexes of these commonly used classification methods on two classical hyperspectral remote sensing sets,the comprehensive evaluation of each algorithm is carried out,which provides a certain basis for the selection of hyperspectral remote sensing image classification methods.At the same time,because the traditional sparse representation model only considers the spectral information of the pixels and does not analyze the spatial relationship of the pixels,which leads to low classification accuracy,this paper constructs a Joint Sparse Representation Classification(JSRC)model,which considers that the pixels of the center to be measured are similar to those of its neighborhood,and takes full account of spatial information.By setting up a neighborhood,the pixels in the neighborhood are regarded as a whole,and the whole is solved.The result is the classification result of the central pixel.The experimental results show that the traditional classification algorithm based on statistical mode has serious salt and pepper phenomena,poor classification effect and low classification accuracy for contours.Accuracy is over 80%,The SVM-based classification method using RBF kernel function can get better classification results,but the RBF kernel function is time-consuming and inefficient in finding the optimal parameters.Accuracy is over 87%,The classification method based on SRC model is more effective than that based on statistics.Accuracy is under 88%,Compared with JSRC model with spatial information,JSRC model still has some shortcomings.With appropriate scale and sparse coefficient,the accuracy of JSRC model is higher than other methods,and the classification effect is better.Accuracy is over 95%,It is a more suitable classification method for hyperspectral remote sensing images.