Scene Classification Of Remote Sensing Image Based On Convolutional Neural Networks

The scene classification of remote sensing image facilitates the interpretation of the semantic content of massive high-resolution remote sensing images,and can provide related information to solve practical problems,such as target identification and target detection.The interpretation quality of remote sensing image is directly determined by the accuracy of classification.In recent years,convolutional neural network is widely used in image classification,because it can extract abstract and high-level features.Aiming at overcoming the difficulties in the field of scene classification of remote sensing image,such as the complexity of artificial feature extraction,difference in the same class and the similarity among different classes of scene,a method based on convolutional neural network is proposed in this paper.And a novel process to solve the difference and similarity of the remote sensing scene images is proposed.Also,two improved models of the network are discussed.The details are as follows:(1)The basic structure and training process of traditional convolutional neural network are mainly stated.According to the specific experimental factors,the structure of the convolutional network,the number of convolutional cores,the parameters in the training period,such as the learning rate and batchsize,are altered to study the effect on the accuracy of scene classification.The suitable convolutional network model and network parameters are selected.The result is compared with the traditional scene classification methods and model with dropout layer and the visualization analysis is also carried out,which proves the feasibility of the traditional convolutional neural network model applied to scene classification of remote sensing images.(2)By using the classification ability of convolutional neural network,specific target from the scene image is extracted and identified,which can provide information to judge the kind of scene it belongs.The target of the ship is extracted and identified in this paper in order to distinguish the scene classes between the ocean,the port and the river.And the extracted vessel type information is used to judge and correct the type of scene.This process suggests a new idea to solve the difficulty of scene classification of remote sensing images.(3)Two improved methods are proposed based on the traditional network model:an improved method is to replace ReLU,the commonly used activation function in convolutional neural network,by PReLU function.Another method is aiming at solving the problem of small data set.The model is pre-trained on a larger data set,and then fine-tuning the model with a specific data set.AlexNet and GoogLeNet network model is selected in this paper,and the remote sensing scene classification data set is used to fine-tune the model.Through the experiments,the two improved methods have higher accuracy than the traditional model.