Study On Land Coverage Classification Of Hyperspectral Image Based On Machine Learning

In recent years, land cover classification of remote sensing hyperspectral data has become an interesting area. The main feature of hyperspectral data is that it integrates the traditional image dimensional and spectral information. It gets the continuous spectrum information for each pixel spatial images and surface information, enabling the feature information retrieval and object recognition based on the spectral features. Hyperspectral data contains manifold structure and spatial information and the labled data is expensive and the number is small. The main work is:Firstly, Laplacian SVM is introduced into the hyperspectral classification. Laplacian item projects the hige dimensional data into a low dimentional space and find out its manifold structure. The pplication of Laplacian SVM in hyperspectral classification make well use of the characteristic of the hyperspectral data to reach a higher accuracy.we introduce Laplacian Pre-conjugate gradient into the hyperspectral classification. The primal and dual optimization are two different ways of solving Laplaician SVM. Pre-conjugate gradient solution of Laplacian SVM solution adds the dual frame gradient-based decline in the early in the original method to predict the stability of the stop condition, which greatly reduces the time and complexity of training.Secondly, in this paper, we proposes the Simulated Annealing Algorithm the classification parameters optimization in different classification machine. The three parameters in the Laplacian SVM includes kernel parameter, the penalty factor, and Laplacian factor. Compared with the traditional grid-searching method, the simulated annealing algorithm to optimize the parameters saves a lot of time, especially if there are many parameters.In this paper, we conducts an experiments on the Indian Pine dataset, trying to classify the six classes. At the same time we compare the proposed method with different classifiers. This experiment shows the high accuracy and lesser time of the method.