Based On Spectral Image Classification Algorithm Of Support Vector Machine And Its Properties

Remote sensing image classification in digital image classification process is a very important research direction which has a lot of things in common with ordinary image classification. According to its own characteristics and problems faced by remote sensing images, if we can make the computer in a meaningful automatic classification of remote sensing images, it will allow people to study can be more convenient to filter and applied in the face of massive remote sensing data, so that it can continue to draw on remote sensing image classification algorithm, the improvement and innovation of great significance.Support vector machines (SVM) based on statistical learning theory is a machine learning algorithm, and the whole experiment section is divided into two parts. It finds optimal separating hyperplane by optimization problem solver in the high dimensional feature space, so as to solve the classification problem of complex data. SVM theory which is applied to the study of remote sensing image classification is still in the development stage. It is starting from the late1990s, SVM is used for classification of remote sensing data. The algorithm is first applied in multi-spectral remote sensing image classification. Because of SVM supporting the classification of high dimensional feature space, a large number of researches began to research for classification of hyperspectral data. The results show that the algorithm is almost free from the "curse of dimensionality". The results show that SVM in high-dimensional feature space can also obtain more high classification accuracy.The six-band multispectral data and50-band hyperspectral data were selected to carry out two series of experiments. The first part is a comparative study on the classification of different kernel functions. Finally we analysis out in hyperspectral image classification process of adjustment parameters, and select the images of the different types of nuclear function of classification whose accuracy is not significant. At the same time using different kernel functions in the classification of multispectral images, the experiment results show that the linear kernel and polynomial kernel function showed better classification accuracy. The article does not get better classification accuracy because of rbf kernel function involving the adjustment parameters. Through the analysis of the penalty parameter, the second part for changing the penalty parameter C, the weight parameters W, the training samples N comparative experimental study, can be learned for adjusting the parameters of kernel function. The punishment is essential of the factors, plays an important role in the classification. From the analysis of sample weights, increase the sample weights not only enhance classification accuracy of the corresponding class samples, but also increase the improvement of overall accuracy due to the overall mechanism of the SVM. So in the test sample classification, due to the number of samples is not balance, you can adjust the weights to enhance the overall classification accuracy. The experimental training samples are added due to Foodys’small sample SVM for multi-spectral classification of remote sensing data. The experiments show that the SVM training sample selection is different from the traditional method of maximum likelihood. By selecting a representative sample of the training data, can also achieve the classification accuracy of the traditional large sample, fewer the number of support vectors and less computational complexity.On the one hand, we use both hyperspectral data and multispectral data classification which is an innovation in the text. Hyperspectral image classification process we did not analyze the specific kernel function, but analysis accuracy when optimal combination of each core function of the number of classification reach the best result. Generally speaking, it has little differences reflected in hyperspectral image classification when various kernel function parameters achieve optimal. In multi-spectral study, we used a specific kernel function parameter values to the value of a set of classification results and classification accuracy. The experiment demonstrates the type of kernel function with spectral image classification accuracy is not the same when we do not select each parameter.On the other hand, from the experiment to analyze the image of SVM kernel function of the spectral image classification, we can get conclusions by training samples from three aspects experimental comparison. SVM classification process characteristics summed up the characteristics of SVM in spectral image classification. It plays the role of reference for future research on nuclear function and related parameters in hyperspectral image classification process through the analysis of experimental results.