Hyperspectral Image Classification Based On Multi-architecture Convolutional Neural Network F A S

In recent years,with the rapid development of hyperspectral remote sensing technology,hyperspectral imaging is poised to enter the mainstream of remote sensing.Hyperspectral remote classification technology has great potential and application value in many practical applications,including resource exploration,agricultural production,medical detection,geological disasters,atmospheric environment,archaeological relics and military reconnaissance.The detailed spectral information provided by hyperspectral sensors improves the capacity to differentiate the interesting land-cover classes.However,the curse of dimensionality occurs caused by high-dimensional spectral bands.It’s a challenging to effectively utilize the spatial and spectral information and extract discriminative features for hyperspectral image classification.Convolutional neural network is introduced into the classification of hyperspectral image by many scholars due to the powerful capability on image classification.However,single fixed-size spatial architecture hinders the excellent performance of CNN due to the neglect of various land-cover distributions in hyperspectral image.Meanwhile,convolutional neural network depend on a large quantity of training data due to the models being heavily parameterized.However,only limited training samples are available in hyperspectral image data.The CNN model tends to be over-fitting for hyperspectral image classification.In this paper,considering characteristics of hyperspectral image data and various land-cover distributions in hyperspectral image,the deep multi-architecture convolutional neural network model is constructed to hierarchically and automatically extract the discriminative deep spatial and spectral joint features for precision hyperspectral images classification.The main works of this paper are as follows:A novel divide-and-conquer dual-architecture convolutional neural network is proposed for hyperspectral image classification.The dual-architecture CNN is designed considering various land-cover distributions of hyperspectral image classification.A new regional division strategy based on local and non-local decisions is devised to divide hyperspectral image into homogeneous and heterogeneous regions,where a multi-scale CNN architecture is constructed to learn joint spectral-spatial features in the homogeneous regions and a fine-grained CNN architecture is constructed to learn hierarchical spectral features in the heterogeneous regions.Experimental results on different hyperspectral image datasets demonstrate that the proposed method provides encouraging classification performance,especially region uniformity and edge preservation.Then,a based on adaptive spatial and spectral constraint data augmentation and divide-and-conquer dual-architecture convolutional neural network is proposed for hyperspectral image classification.A novel data augmentation method based on spectral similarity under adaptive spatial constraint is devised to alleviate the over-fitting problem.Experimental results demonstrate the effective of data augmentation,the robust and the advancement of the proposed algorithm on different hyperspectral datasets.Finally,a based on adaptive diverse region ensemble convolutional neural network hyperspectral image classification method is proposed.This method uses multi-region convolutional neural network to extract diverse spatial features and utilize adaptive weights to learn regional features which are more important for classification.The model pays more attention to the input regions that are related to the central pixels and are advantageous for classification,and weaken the influence of regions which are conducive to central pixels and are not disadvantageous to the classification.The diverse-region ensemble loss function is designed to cooperatively optimize the multi-region convolutional neural network branch and regularize the extracted features of each branch network for obtaining better classification results.The experimental results on three hyperspectral image data show that the proposed algorithm is superior to other advanced classification algorithms,especially region uniformity and edge preservation.