Research On Remote Sensing Image Classification Based On Deep Convolutional Neural Network

The accuracy and recognition time of remote sensing image classification are important indicators for verifying network performance.With the continuous development of remote sensing satellite technology,Deep Convolution Neural Network（DCNN）is more efficient in the field of remote sensing image processing,abstracting the original image and obtaining high-level semantics Features closer to the object overcome the disadvantage of traditional remote sensing image classification means.To improve the performance of DCNN networks,this paper proposes two different ways to improve it based on a detailed study of DCNN theory.The research content completed in the paper is as follows:First of all,in view of the problem that the performance of deep convolutional neural networks is mostly affected by the activation functions,this paper combines the characteristics of the logarithmic function and the arctangent function to propose a new activation function sArcReLU.The negative semi-axis of this function uses the characteristics of the arctangent function to solve the problem of "neuron death" caused by the constant zero output of the negative axis of the ReLU function.The variable positive semi-axis can continuously correct the sparse distribution of data and eliminating the positive of the ReLU function.The linear characteristic of the semi-axis solves the phenomenon that gradient disappears in training network model.The experimental results show that compared with the ReLU activation function and ArcReLU activation function,the sArcReLU function can more effectively advance the classification performance of the network model.Then,an image classification method based on the fusion residual module is proposed.A convolutional neural network is designed in a targeted way,the residual unit is introduced on the third and fourth convolutional layers of the network structure,Incorporate Dropout regularization method and the softmax classifier is used for classification.Compared with the original network model on two remote sensing data sets,NWPU VHR-10 and UCM,the experimental results show that the algorithm in this chapter has higher classification accuracy.Finally,the improved model proposed in this paper is actually applied to the new dataset RSI-15,and the accuracy of actual remote sensing image classification verifies its network performance.The experimental results show that the two network models proposed in this paper have good application foreground.