Research On Scene Classification Of High-resolution Remote Sensing Image Based On Convolutional Neural Network

With the rapid development of high-resolution remote sensing technology,the amount of remote sensing image data has explosive growth,and massive image data urgently needs to be classified.However,traditional classification methods can no longer meet the needs of production and life in terms of classification effect and efficiency.Therefore,a scene classification method has become one of the research hotspots in the field of remote sensing image classification.Convolutional neural networks(CNNs)have made breakthrough progress in the field of computer vision.New ideas and innovation have been emerging.The basic idea is to extract some edge features at the bottom layer and get the main features in the deep layer of network.Gradually,the semantic information is described from the low-level to the high-level.According to working principles and mechanisms,convolutional neural networks are very suitable for scene classification tasks.In this thesis,theory and practical experience of convolution neural network in the field of computer vision are applied to high-resolution remote sensing image scene classification.Convolution neural networks have strong ability of feature learning and expression,which can reduce the dependence of classification work on manual work and liberate the workers from the heavy task of classification.The main work of this thesis is summarized as follows:(1)The basic concept and theories are first briefly introduced,including the structure of convolutional neural network,the process of CNN training,the principle of attention mechanism,the structure of residual network,loss function.Also,it is described that the latest remote sensing image scene classification dataset.(2)An improved softmax cross-entropy loss function is proposed,called S-softmax loss function.This method makes up for the defects of the intra-class variance in the traditional softmax.The main idea is to set the hyperparameter s to increase the inter-class spacing.(3)An improved model,Aresnet,is proposed.This model integrates Res Net50 with Convolutional Block Attention Module(CBAM).The key of scene classification is to extract key information effectively and ignore the interference of irrelevant information,Aresnet integrated with CBAM can pay more attention to the feature areas,and then increase the accuracy of classification.The large amount of experiments have been conducted in order to evaluate the proposed methods by using the NWPU-RESISC45 dataset and AID dataset.The experimental method is to set different conditions(such as CBAM module,loss function),record and analyze the experimental results under different conditions and different datasets.The experimental results show that the Aresnet model applied with proposed in this thesis has achieved a classification accuracy of 96.4% on the NWPU-RESISC45 data set and 97.3% on the AID data set.