Hyperspectral Imagery Classification Based On Rotation Deep Forest And Deep Filtering

With the rapid development of hyperspectral remote sensing technology,hyperspectral remote sensing plays an increasingly important role in the field of remote sensing,and has been widely applied in many fields,such as geological exploration,precision agriculture,environmental monitoring and so on.Hyperspectral remote sensing data are characterized by many wavelengths,high spectral resolution and high spatial resolution.When hyperspectral images are analyzed and processed,rich spectral information and spatial information can be used.Extracting effective features from hyperspectral data and classifying them with spectral and spatial information are the most common tasks in hyperspectral research.In this paper,hyperspectral data is classified by rotation-based deep forest,deep edge preservation filtering and spectral-spatial classification based on random patch convolution feature extraction.The spatial information and output probability vector information are used to further improve the classification effect and efficiency of hyperspectral data.The research contents of this paper mainly include the following aspects:1.In the field of remote sensing,the convolutional neural network method is used to deal with many challenging tasks,including hyperspectral image classification.However,the training of deep model is time-consuming and requires a large number of label samples.The classification performance of deep models is limited due to the small number of labeled samples in hyperspectral images.In this paper,a rotation-based deep forest classification method is proposed.The rotation forest is used to transform the hyperspectral image to generate new features.The idea of multi-layer structure is combined with the new feature and the probability vector of each layer as the input of the next layer,and the adjacent pixels are added to each layer to add spatial information of the data.The method subtly combines the spatial features of the hyperspectral image and the vector features of the output probability,and achieves a good classification effect.2.The output probability vector is filtered by edge-preserving filtering,which can effectively remove the noise of the probability vector and improve the classification effect.The deep edge preserving filter proposed in this paper is a new hyperspectral image classification algorithm based on filtering.Firstly,the classification probability result graph is obtained by using the radial basis function-based SVM to classify the hyperspectral image,edgepreserving filtering is performed for each dimension of the classification probability result graph.Then,the deep edge-preserving filter model adopts the idea of multi-layer structure,and each layer uses the radial basis function-based SVM as the classifier,the classification probability features of filtered and the original features are combined into the input of the next layer.After filtering,the output probability vector not only improves the classification effect,but also enriches the characteristics of the samples,which conbined with the original feature combination will further improve the classification accuracy of hyperspectral.3.Hyperspectral images contain abundant spatial information.Random selection of a certain number of random patches from hyperspectral images and convolution of them can effectively extract the spatial features of hyperspectral images.This paper proposes a new spectral-spatial classification method based on random patches convolution feature extraction.Random patches are selected from the reduced hyperspectral image for convolution.The extracted spatial features are combined with spectral features and introduced into adjacent pixels,and the classification effect is improved by using spatial information again.