Hyperspectral Image Classification Based On Feature Learning And Deep Learning

Hyperspectral image technology has transformed from a scarce research effort into a wide range of community available products for users over the past four decades.Especially in the past 10 years,a large number of methods for hyperspectral data processing have emerged according to the characteristics of hyperspectral imagery.The classification of hyperspectral images has becoming a hot topic,which has attracted countless researchers.The traditional classification method only considers the spectral information of the hyperspectral image and ignores the spatial information,resulting in a poor classification result.The classification results can significantly improved after combining spectral information and spatial information of hyperspectral images.However,the methods for extracting spatial information are very different from each other.How to extract the spatial information of hyperspectral images effectively and remove the redundancy of hyperspectral high-dimensional data becomes the key to the study.In recent years,with the development of artificial intelligence,deep learning technology has gradually entered into people's field of vision.Deep learning firstly came out in image recognition,but in a few short years,it spread to all areas of machine learning.Today,deep learning has a very good performance in many machine learning areas.This paper takes use of the advantages of deep learning in feature extraction to apply deep learning to the task of classification of hyperspectral images.The main contributions are as follows: 1)Aiming at the defect that the traditional method does not combine spatial information of hyperspectral images,a hyperspectral image classification method based on feature fusion and deep forest is proposed.The method first uses PCA to reduce the original hyperspectral data,and then extracts its extended morphology profile(EMP),linear multiscale features,and nonlinear multiscale features on the reduced-dimensional data.The features merged by these three features are then send to the deep forest classifier.The experimental results show that the hyperspectral classification method based on feature fusion and deep forests is highly efficient,and has excellent performance even on small samples.2)For the problem that the fixed geometric structure of CNN is not good enough for transforming the geometry of the model,which results in the problem that CNN may not be recognizable after reversing an image.A hyperspectral image classification method based on deformable convolutional neural network is proposed.This method introduces a new module named deformable convolution on the basis,which can identify images with different shooting angles,and enhance the ability of the traditional CNN to transform the modeling geometry.The experimental results show that the deformable convolutional neural network has a very good performance in the hyperspectral classification work.3)The internal data representation for CNN does not consider the important spatial levels between simple and complex objects,which makes it impossible for CNN to identify photos taken at different angles accurately.A hyperspectral image classification method based on capsnet was proposed.The capsule network simulates that the representation of objects in the human brain does not depend on the angle of view,and combines the relative relationships between objects to improve the recognition rate of objects.The method firstly uses PCA to reduce the dimension of hyperspectral image spectral dimension and remove redundant information.Then use window scanning to obtain training and test samples.The experimental results show that the capsule network also performs well in the task of classification of hyperspectral remote sensing images.