Hyperspectral Image Cross Domain Classification Based On Adaptive Residual 3D Convolutional Neural Networks

Hyperspectral imagery has become an important resource in the field of remote sensing data analysis because it has both spatial and spectral rich information of the target scene.It has been widely used in precision agriculture,environmental monitoring and military fields.Feature extraction and classification technology based on hyperspectral images are the basic problems of hyperspectral-image analysis,and is also the basis of automated agriculture and precision agriculture.And its accuracy of classification will directly affect the performance of advanced agricultural visual tasks.The existing feature extraction methods mainly use linear or non-linear equations to artificially design and extract specific features.These processes not only require significant professional knowledge and experience,but also call for a lot of time.Additionally,it makes the artificial feature extraction method Inefficient that the hyperspectral images have a great number of channels,wide band coverage and scarce samples,resulting in unsatisfactory classification results.How to classify ground objects accurately and stably is still an important problem to be solved urgently.In order to solve these problems,this paper studies the above problems from three aspects based on deep learning method which has developed rapidly in image processing field in recent years,which are feature extraction of hyperspectral images,network model optimization and parameter optimization,feature utilization and classification.The main innovative work of this paper is as follows:Firstly,this paper studies suitable the feature extraction method for hyperspectral image classification based on the deep learning method.According to the characteristics of large amount of spectral information and large number of channels in hyperspectral images,a residual three-dimensional convolutional neural network is designed.The three-dimensional convolution kernels are used to extract spatial and spectral information simultaneously,which overcomes the redundancy problem of image data and network parameter in two-dimensional convolutional neural network for hyperspectral image processing.The deep network we proposed realizes spectral-spatial feature extraction and accurate classification,by introducing residual bottleneck structure and normalization algorithm.Secondly,aiming at the hyper-parameter optimization of deep learning network,this paper studies manual method and adaptive TPE method.As hyper-parameters can not be adjusted directly by sample learning,it often needs rich training experience and large number of experiments to optimize.In this paper,by setting the hyper-parameter search space on the artificial preset network,we adjust the hyper-parameters adaptively.The quantity of network parameters and training time are effectively squeezed without complexity of violent searches.Finally,in the terms of feature utilization and classification,this paper studies the cross-domain classification of farmland hyperspectral images.By utilizing the advantages of three-dimensional convolution kernel and migration training methods,we can realize the goal of training shared parameters in a dataset with more training samples,and classifying in a dataset with fewer labeled training samples.Moreover,even if there are totally different types of objects between the two datasets.It can be classified accurately.In summary,this paper studies the hyperspectral image processing method based on deep learning methods and characteristics of hyperspectral data structure,and makes a useful exploration on the application of the proposed method in hyperspectral cross domain classification,which has certain theoretical significance and application value.