Research On Hyperspectral Image Classification Based On Improved Convolutional Neural Network

With the increase of the spatial resolution of hyperspectral remote sensing images and spectral dimensions,how to fully extract the deeper feature information of hyperspectral images has become a research hotspot in the classification of hyperspectral images.Deep learning has efficient feature extraction capabilities and classification performance.As one of the most important models in deep learning,convolutional neural network(CNN)has been widely used in image processing.In the processing of hyperspectral image classification,convolutional neural network has been proved to be a good classification model.However,in the local feature learning and deep feature extraction of hyperspectral images,the performance of the traditional convolutional neural network model structure is not satisfactory.In this thesis the improved convolutional neural network model is applied to the classification of hyperspectral images,and the following main research works are done to strengthen the CNN model for local feature learning and deep feature extraction of hyperspectral images:(1)To improve the feature extraction ability of the CNN model,this thesis designs the network model with three modules: shallow feature extraction module,attention-weighted learning module and deep feature extraction module.Among them,the shallow feature extraction module is used to learn the shallow features of the image and the attention mechanism is introduced into the attention-weighted learning module to perform weighted learning of the local key features of the image.The deep feature extraction module is composed of several residual blocks,which is used to learn the deep features of the image and avoid the problem of network degradation while increasing the depth of the network.(2)In order to make full use of the rich spatial spectrum information of hyperspectral images,three 2D-CNN network models are designed with spectral,spatial and spectral-spatial information.In addition,attention mechanism and residual network are introduced into the three models.Experi-mental comparison verifies the feasibility of improving the CNN model based on the attention mechanism and residual block.(3)To maintain the three-dimensional characteristics of hyperspectral image data,the improved hyperspectral image classification method under the 3D-CNN model is realized without increasing the network parameters,it is proposed to use the dilated convolution to replace the ordinary convolution in the residual block,thereby expanding the receptive field of the convolution kernels,so that the model can learn more feature information of the image.Finally,the experimental results show that the improved network model has achieved better results in hyperspectral image classification.In summary,in order to better extract local features and deep features of hyperspectral images,the attention mechanism,residual block and dilated convolution structure are introduced into the CNN model respectively.Experimental results verify the effectiveness of the improved CNN model in hyperspectral image classification.