Classification Of Hyperspectral Remote Sensing Image Based On Convolutional Neural Network

Hyperspectral remote sensing image classification is not only a basic task of remote sensing image automatic understanding,but also a prerequisite for deep mining of remote sensing information.It has a wide range of applications in natural disaster detection,geographic image retrieval,environmental monitoring and other fields.With the development of remote sensing satellite technology,the resolution of remote sensing images obtained is getting higher.It is precisely because the resolution of remote sensing images continues to increase,more useful information can be extracted from the images.However,due to the improvement of the resolution of hyperspectral remote sensing images,the complex spatial structure and information redundancy of the images bring new challenges to the scene classification task: There are not only some same land cover types and object categories among different scenes,but also the objects always appear in different scales and directions.These factors increase the intra-class differences and inter-class similarities,which require higher requirements for the classification model.In recent years,compared with traditional classification methods,convolutional neural networks have made significant progress in remote sensing image classification tasks,and have become the mainstream method for studying hyperspectral remote sensing images.Although the convolutional neural network brings a lot of convenience to the processing of high-resolution remote sensing images,it is hindered by its high computational cost in many practical application deployments,making the compression of the network model attract more attention.Based on this idea,this thesis introduces pruning operation to convolutional neural network,in order to compensate for the precision loss caused by model pruning,the convolutional neural network was combined with knowledge transfer technology,this thesis proposes a transfer learning method based on soft target and activation feature map.These methods can effectively compress the convolutional neural network model and ensure the accuracy of the model.The main works of this paper are summarized as follows:1)This thesis proposes a hyperspectral remote sensing image classification algorithm based on soft target knowledge distillation.our method is on the basis of the theory of soft target knowledge transfer of convolutional neural network.The attention mechanism is introduced into the network structure to strengthen the extraction of important features.The network performs model pruning,and then uses soft target knowledge distillation technology to transfer learning to the model to compensate for the loss of classification accuracy after model pruning.The experimental results show that the proposed method obtains good experimental results on the three commonly used hyperspectral remote sensing datasets.2)This thesis proposes a hyperspectral remote sensing image classification algorithm based on activated feature map.The algorithm pretrains a teacher network with a deep structure,using the activation feature map of the teacher network to guide the pruned student-network,and realizing the knowledge transfer of teacher-network.Experimental results show that our method can achieve good classification performance on hyperspectral remote sensing datasets,and it also be combined with the above soft target knowledge distillation method to train the student network,which will further improve the classification accuracy of the network.