Hyperspectral Classification Method And System Based On Multiscale Spatial Spectrum Combined Convolution Neural Network

In recent years,due to high spectrum images contain rich space and spectral information,its processing and analysis have been widely concerned,including classification,unambiguous,variation detection,and target detection.With the development of deep learning technology,high-spectral image features can be more effective due to depth learning methods,making great attention based on high-spectral classification methods based on deep neural network.Due to the high spectral image spectral resolution,the characteristics of more spectral bands and mixed large amounts of mixed beam,facing computational complexity,time consuming,and existence of large amounts of redundant information when performing high-spectral image classification.Difficulties.How to fully integrate the spectral information and spatial information of the high-spectral image under a limited training sample,and quickly and stable classified high-spectral images have become a huge challenge.In this paper,a high-spectral classification model based on multi-scale empty-spectral co-convolutional neural network is proposed in combination with the spectral space dual-branch multi-scale model.Experiments in real high-spectral images verify the effectiveness of this classification method and system.The main tasks of the paper include:（1）High Spectrum Image Classification Model and Algorithm for Convolutional Neural Network Based on Dual Branch Multi-Scale Empty-Spectrum.In this paper,based on the high-spectral image classification method based on deep neural network,the double branch structure based on convolutional neural network is extracted,and the spatial features are extracted,and different scales of the image are combined with different sizes.And gradually fuse the information under different scales through the full connecting layer at different stages.In addition,the use of a dense connection reduces the excessive problems due to small training samples.By experiments with multiple real high-spectral images,and some classical algorithms such as SVM and CDCNN,and with unbound branches and multi-scale models,this paper proposed double-branch multi-scale empty-spectral classification The model can effectively improve the classification effect.（2）In order to increase the weight of different branch spectroscopy and spatial information,the spatial spectrum feature information of different branches is fully excavated,and the high-spectral image classification accuracy is further improved,and a high spectrum of convolutional neural networks of double branch dual focus modules is proposed.Image classification model and algorithm.In order to further improve the representation of effective information,this paper is based on the channel relationship of the double-branch convolutional neurocal network spatial spectrum branch,and the binding attention to the binding mechanism is used to consolidate the nerve model,which is attached to the double branch.The method of two parallel attention modules,captures effective spatial spectroscopy information,and finally blends the output of the two attention modules,further enhances the characteristics of the high-spectral image.The effectiveness of the classification method proposed in this paper is shown by experiments on a real high-spectral image.（3）On the basis of the above model and algorithm,a high-spectral remote sensing image classification software system is implemented based on the Python GUI frame design,mainly including high-spectral image management,model selection,high-spectral image classification,and analysis and other modules.The article gives the main process of the entire system framework and the development implementation and testing of the core module.