Research Of Remote Sensing Image Scene Classification Algorithm Base On Convolutional Neural Network

With the continuous development of remote sensing observation technology,the number of remote sensing satellites in multiple series has been increasing,and more and more remote sensing images of different scenes have been acquired.Because remote sensing images not only have rich spatial and texture features,but also contain a large amount of scene semantic information.Therefore,remote sensing image scene classification has always been an active research topic in the field of remote sensing image analysis.In recent years,remote sensing image scene classification has been widely used in urban planning,geographic image retrieval,natural disaster detection,environmental monitoring,vegetation mapping,and geospatial target detection.Traditional remote sensing image scene classification algorithms mainly rely on manual features designed by professionals.Most of them acquire simple physical features such as edges,textures,and colors.The classification accuracy is low and cannot meet the requirements of practical applications.Compared with traditional algorithms,convolutional neural networks can extract abstract high level semantic features from images,and have powerful feature extraction capabilities and generalization capabilities.Therefore,this thesis is based on convolutional neural networks.For the complex background of remote sensing image,some scenes with many small and large scale targets,large differences within scenes and high similarity between classes.Two remote sensing image scene classification methods are proposed:multi branch feature fusion and dual attention.The specific work content is as follows:(1)In order to enhance the model's ability to extract scene features,a deep multi branch feature fusion network is designed with the residual network as the backbone network.Since features at different levels contain different types of information,a multi branch convolutional neural network is proposed to extract low level,middle level and high level feature information from remote sensing images.These three features are fused to reflect the scene information more comprehensively.When fusing three different features,in order to reduce the complexity of feature fusion,a feature grouping fusion method is proposed.Feature grouping fusion method divide the three levels of features extracted by the multi branch network into several groups,and fusion them in each group respectively,so that the features in each group include three levels of features,and then aggregate multiple groups and output.Experiments are carried out on three public data sets of UCM,AID and OPTIMAL.The experimental results prove the effectiveness of the proposed deep multi branch feature fusion network.(2)For the cross-entropy loss function,the same weight is assigned to each sample,and the difficult-to-classify and easy-to-classify samples cannot be treated differently.The modulation factor is introduced and added to the cross-entropy loss function to reduce the loss of easy-to-classify samples and pay more attention to difficult-to-classify samples.And for the cross-entropy loss function,only the loss of the correct label position is considered,and the loss of the wrong label position is ignored.Introduce a fixed distribution,taking into account the loss of correct and wrong label positions,to promote better learning of the model,close to the correct classification,and as far away as possible from the misclassification.The two parts for the improvement of cross-entropy loss are combined and balanced by the hyperparameter ? to form the difficult sample label position loss function.On three public data sets,through multiple sets of comparative experiments,the feasibility of the proposed loss function is verified.(3)In order to pay more attention to salient regions and salient features in remote sensing image scene,a dual attention dense network is designed.An adaptive spatial attention module and an adaptive channel attention module are designed respectively.In the adaptive spatial attention module,the adaptive parameter activation function is introduced to assign different nonlinear transformations to the input features in the spatial attention network.The spatial attention network aims to generate the attention map based on the local features,which contains the important weights of different regions.Finally,the attention map is applied to the global feature to realize the function of paying attention to the salient regions.Adaptive channel attention learns channel attention by capturing adaptive cross channel interaction range,and generates important weights of each channel.The adaptive parameter activation function is introduced to adjust the feature values of different channels.Finally,it works with the global features to attention the salient features.Experiments are carried out on UCM,AID and OPTIMAL data sets,and compared with various algorithms,the effectiveness of the dual attention model is proved.