Research On Scene Classification Of Remote Sensing Images Based On Convolutional Neural Network

The remote sensing image contains rich information,which can truly reflect the surface coverage.It is an important basis for the study of ecological environment change,land resource management and sustainable development.In many application fields,it is the most basic and extensive application to obtain the surface coverage by classifying the remote sensing images,which attracts more and more researchers' attention.Because the traditional "pixel oriented" and "Object-Oriented" classification methods can not meet the needs of high-level semantic classification tasks of remote sensing images,the "scene oriented" classification method has gradually become a hot research topic in the field of remote sensing image classification.Because of the great success of convolutional neural network(CNN)in natural image classification,target detection and other fields,the application of convolutional neural network in remote sensing image scene classification has become a research hotspot in recent years.However,there are still some problems in convolution neural network,such as over fitting,local optimal solution based on gradient descent,and time-consuming manual adjustment of learning rate parameters.The main contents of this paper are as follows:(1)In order to solve the problems of over training in the whole connection layer of convolutional neural network,such as poor generalization performance,low classification accuracy and over fitting risk,this paper proposes an improved strategy to reduce the structural complexity of convolutional neural network model to mitigate the risk of over fitting.The convolution neural network is used as feature extractor and the extreme learning machine(ELM)is used as classifier to construct a hybrid classification model for remote sensing image scene classification.The experimental results show that this method effectively suppresses the risk of over fitting and improves the classification accuracy.(2)In order to solve the problem that the optimization algorithm of convolutional gradient descent(SGD)is easy to fall into the local optimal solution,and manual parameter adjustment is too time-consuming and inefficient,this paper proposes a strategy of combining the whole local optimization algorithm and SGD algorithm by using evolutionary strategy algorithm.In the process of optimization,evolutionary strategy algorithm and SGD algorithm alternate.In the SGD step,annealing strategy is used to make the model automatically adjust the learning rate,solve the problem of too time-consuming and low efficiency of manual parameter adjustment;in the evolutionary step,elite strategy is used to make the model obtain better global optimal solution.In the case of setting the optimal parameters and hidden layer nodes,a convolutional neural network model with this strategy is constructed to study the scene classification of remote sensing images.Experimental results show that the strategy can effectively avoid falling into the local optimal solution andrealize the automatic adjustment of learning rate parameters,save time and improve efficiency,and effectively improve classification accuracy and consistency.