The Research On Multi-scale Hyperspectral Image Classification Based On Deep Convolution

Hyperspectral remote sensing technology is a method that combines the unique spectral features of ground objects with the spatial features that reflecting the shape texture and layout of ground objects to achieve accurate detection,recognition and attribute analysis of ground objects.Hyperspectral images obtained by hyperspectral remote sensing technology not only contain spectral information reflecting the unique physical properties of ground objects,but also provide a large amount of spatial information.By making full use of spectral and spatial information in hyperspectral images,we can accurately identify and classify ground objects,which are widely used in precision agriculture,environmental protection,geological survey and many other fields."Nonlinearity","foreign matter with the same spectrum" and "integration of maps" are the main characteristics of hyperspectral images.At the same time,it contains abundant spectral and spatial information,and has great application potential.With the continuous research and application,the data itself has many problems,such as high-dimensional characteristics,data redundancy,few trainable samples,large module parameters and difficult to extract image features,which increase the difficulty of hyperspectral image classification.In order to solve the above problems,this paper has carried out correlation analysis and research,and proposed two hyperspectral image classification models.The main work is as follows:(1)Aiming at the problems of high dimension,high redundancy,large number of module parameters and difficulty in extracting image features of hyperspectral image data,this paper proposes an improved depth convolution network model with adaptive channel attention mechanism.The incremental principal component analysis and small batch K-means are used to reduce the dimension of the original image,and the depth three-dimensional convolution module and the depth two-dimensional convolution module are constructed to extract the deep spatial and spectral features in the image,as well as the local spatial and edge features around the image respectively.Dual-channel structure is used to extract multi-scale features of images.According to different extracted features,combined with improved adaptive channel attention mechanism,global features,local spatial features and edge features are enhanced,and the enhanced features are fused and classified at the classification level.Experiments on hyperspectral image data sets in Indian Pines,Pavia University and Salinas prove that the proposed classification method effectively improves the classification accuracy.(2)In order to extract more comprehensive and finer image features with less trainable samples,this paper proposes a dense connection neural network model based on improved wavelet decomposition.The original hyperspectral image is decomposed by improved four-level wavelet transform,and the global and local detail features are further added by multi-scale analysis,so as to obtain high-frequency and low-frequency dimension reduction components.The spatial and frequency domain features are deeply extracted and enhanced by 3D feature extraction and enhancement module,and the output features are reused in channel dimension by improved dense connection module,which enriches the feature layer,makes up for the lost features and alleviates the gradient disappearance.A smaller number of training samples(Indian Pines,Pavia University,Salinas)are selected for experiments,and the classification method obtains better classification results.