The Study And Application About Remote Sensing Image Classification Algorithm Based On Dictionary Learning Sparse Representation

With more and more high spatial resolution remote sensing data come to our life,high precision of remote sensing data classification is always a difficult problem to along-term remote sensing science. Especially in the aspects about how to takeadvantage and improve the existing abstract technology, integrate the multi-scaleinformation extraction in order to take the classification and the location informationmore convenient and accurate. They are all scientific problems which worthy ofattention and deep research.In order to solve the low accuracy problem about multispectral remote sensingimage classification, this text set from two kinds of schemes to achieve the effects.First plan, for the classification of the image not only uses the spectral characteristics,but also with the transformation characteristics named normalized differencevegetation index and tasseled cap transformation. We also apply the dictionarylearning sparse representation algorithm to remote sensing image classification. Theresults show that no matter on the classification accuracy and visual effects, thismethod has a certain advantage; Second plan, a new method based on gray levelco-occurrence matrix (GLCM) texture features is proposed. Through analyzing thecorrelation coefficients of the six bands with each other of remote sensing images,two of the most irrelevant bands are selected to extract the texture features of differentmeasurements and combine the features of original bands as the final classifiedfeatures. The classification is based on the study of sparse representation with alearning dictionary method. The results show the rationality and effectiveness of thismethod.The Statistic classification of hyperspectral data is challenging because of itshigh dimension and huge amount of data, especially when the labeled trainingsamples are relatively small because of its difficult and expensive to be acquired inmost situations. First we use hyperspectral data with75%,50%,10%and5%amount of training samples compared with SVM, HSVM and sparse representation methods.Then update the dictionary gradually to get the optimal atoms. We also use the Lassoalgorithm to conduct the experiment. Thus can get a high accuracy and save time.Based on the sparse representation of remote sensing image processing, it is akind of innovative applications. It not only suitable for the promotion of sparsedecomposition but verified its advantages in remote sensing image classification. Atthe same time solve the problem of remote sensing image feature extractions andmake efforts in further remote sensing image development.