Classification Of Multi-feature Hyperspectral Remote Sensing Image Based On Depth Learning

Hyperspectral remote sensing image has a strong ability to express information,and its spectral resolution is usually in the order of 10 nm.With more application of hyperspectral sensor precision,as well as the development of remote sensing technology for decades,hyperspectral image technique has greatly improved,become extremely valuable tool for the monitoring of the earth's surface,and has a wide application,it has outstanding performance in agriculture,astronomy,environmental science and other fields.The rich spectral dimension of hyperspectral remote sensing image makes it not suitable for the statistical classifier with superior effect on multi spectral images.The hyperspectral remote sensing image contains abundant spatial feature information,how to use the spatial information of the hyperspectral remote sensing image effectively is also a problem.In this paper,hyperspectral remote sensing images are classified from two aspects: the spatial and spectral features of hyperspectral remote sensing images.The main research results are as follows:First of all,using the principal component analysis dimension reduction technique,the spectral dimension is significantly reduced at the expense of the minimal loss of spectral information.PCA is a nonlinear dimensionality reduction technique,which has an ideal effect when dealing with non Gauss distribution data.Secondly,in the hyperspectral image classification algorithm in general,often just study the spectral information of multispectral images,ignoring the spatial information of object space,according to the characteristics of hyperspectral remote sensing image,using the lead filtering method,pixel selection of different size window,feature extraction of hyperspectral spatial multi-scale.By adjusting the weight of each scale feature,the high spatial feature is further optimized.The hyperspectral remote sensing image spectral feature vector and spatial feature vector are combined into the "Spectral-Space" feature vector to extract the feature information of the original hyperspectral image to a greater extent.Then,the advantages and disadvantages of the traditional machine learning classifier based on statistical methods and the application of the new depth learning classifier in hyperspectral remote sensing image classification are studied.By constructing self encoding neural network based on sparse stack unsupervised,the hyperspectral remote sensing training pattern.Stack the autoencoder neural network is a fully connected with the stack from encoder to achieve a depth of neural networks,due to the characteristics of hyperspectral data for parameter adjustment,the application is rarely used in hyperspectral remote sensing image classification.Finally,in the unsupervised neural network depth,add a layer for fine-tuning of the supervised logistic regression,artificial markers by using the image data,the depth of supervised neural network hidden layer parameter fine-tuning,fitting classification tasks to achieve.To improve the accuracy of hyperspectral remote sensing image classification algorithm.