Remotely Sensed Images Classification Based On The Joint Feature And Deep Learning

With the rapid development of Optics and Electronics in recent years,hyperspectral sensors on satellite and UAV can obtain more precise spectral information and higher spatial resolution.The detailed spectral information and spatial information provided by hyperspectral sensors can significantly improve the accuracy of ground targets classification.Therefore,hyperspectral image has been applied widely in many fields,such as Ecological Science,Geological Science,Hydrology and Precision Agriculture.However,there are still many difficulties in hyperspectral image classification,such as dimensional disaster caused by the increasement of spectral dimension,“different body with same spectrum” and “same body with different spectrum” caused by different atmospheric scattering conditions and variability of cloud clusters,etc.At present,with the development of hardware and the emergence of large-scale data sets,deep learning technology has shown good performance in the fields of image classification,natural language processing and remote sensing.It can not only extract different levels of features from the input image,but also effectively fuse these shallow features to generate more abstract and powerful advanced features that can represent the image itself.Therefore,based on deep learning to extract the joint features of spatial domain and spectral domain in hyperspectral images,more efficient hyperspectral image classification algorithms are proposed in this paper.(1)A hyperspectral image classification algorithm based on CondenseNet is proposed.By introducing learned group convolution,a sparse network can be learned automatically during training.The joint features of spatial domain and spectral domain can be extracted efficiently with Light-weight network.A hyperspectral image classification algorithm based on multi-scale dense convolution neural network is proposed.This method introduces two characteristics of multi-scale joint feature and anytime prediction.Therefore,different network structures can be selected according to the difficulty of classification of hyperspectral image features.It can be seen from the results that these two methods can not only improve the classification accuracy of hyperspectral images,but also improve the classification effect of homogeneous region and edge region.(2)A hyperspectral image classification method based on non-subsampled directionlets transform(NSDT)and recurrent neural network(RNN)is proposed.In this method,firstly,NSDT is used to extract abundant spatial and spectral information,and then RNN is constructed to efficiently learn the correlation between the spectra and the variation of spectral band,so that the spatial domain and spectral domain information in hyperspectral image are effectively combined,and a good classification effect is obtained.(3)A hyperspectral image classification algorithm based on extend morphology profile(EMP)feature extraction and stacked convolutional auto-encoders(SCAE).Firstly,EMP based on guided filter is used to extract spatial information of hyperspectral images.Then SCAE is introduced to solve the problem that the stacked auto-encoder ignores the two-dimensional characteristics of the image.The unsupervised method is used to compress the spatial and spectral features.Finally,support vector machine(SVM)is used as classifier.It can be seen from the classification results of each dataset that this method has achieved good classification results for detail region,homogeneous region and edge region.