Research On High-Resolution Optical Remote Sensing Image Scene Classification

As one of the important tasks of high-resolution remote sensing image scene understanding,remote sensing image scene classification has received much attention and research.Different from pixel-oriented and object-oriented image classification tasks,remote sensing image scene classification aims to output higher-level semantic interpretation results by analyzing the real content information of a given scene image.For example,to classify the commercial area,scientific and education area,and industrial area of a given urban region.However,how to extract high-level semantic description based on the underlying visual information of remote sensing scene images,that is,to cross the“semantic gap”,is a key problem faced in the current research on remote sensing image scene classification.In traditional remote sensing image scene classification methods,most of them are based on artificially designed features to achieve semantic information acquisition,but excellent artificial features need to rely on expert knowledge and data priors.As an effective automatic feature learning algorithm,deep convolutional neural networks(DCNN)can effectively extract semantic-level feature descriptions,and have achieved better performance than traditional methods in various computer vision tasks,and have also been successfully applied to the research of remote sensing image scene classification.However,the problems of background information interference,multi-scale features’ semantic ambiguity,which leading to insufficient discriminative description,besides,intra-class diversity and inter-class similarity,and limited training samples available about complex high-resolution remote sensing scene images have not been well solved.To solve the above problems,this paper carried out a series of researchers based on deep convolutional neural networks for high-resolution remote sensing image scene classification.The main research contents are as follows:(1)To improve the discriminative representation ability of fused features and reduce information redundancy,background noise,and semantic ambiguity between multi-scale features,a remote sensing image scene classification method based on adaptive multilayer feature fusion strategy is proposed.This method designs an adaptive feature selection mechanism,which further improves the discriminative ability of fusion features by selecting valuable information in different feature layers.Furthermore,a multi-scale semantic feature aggregation method for remote sensing image scene classification is proposed.This method utilizes attention mechanism and combines multiple deep feature maps to extract relatively accurate and complete discriminative semantic regions.Then,using the learned discriminative semantic regions to guide the extraction and aggregation of multi-scale semantic features.Due to effectively reducing background interference and semantic ambiguity between multi-scale deep features,the proposed method can significantly improve scene classification performance.(2)In order to effectively use the long-range dependencies of deep feature maps to enhance the discriminative representation ability of the given remote sensing images and reduce the computational complexity at the same time,a non-local attention enhancement based remote sensing image scene classification method is proposed.This method improves the classic non-local attention block,and proposes a new cross-layer non-local attention calculation block.Then,a cross-layer non-local neural network for remote sensing image scene classification is proposed.Because this method can aggregate context dependencies between different convolutional feature maps,it has the ability to realize feature enhancement of key local regions and improves the discriminant of deep feature representation.Finally,this method reduces the computational complexity while achieving higher scene classification performance.(3)To effectively solve the problem of intra-class difference and inter-class similarity in high-resolution remote sensing image scene classification,proposing a remote sensing image scene classification method which combines with attention center loss.In the feature extraction stage,a mixed-order attention feature extraction network is designed to capture stronger semantic descriptions by making full use of the first-order and secondorder information of deep features.In the loss function stage,on the basis of the classic center loss,a new attention center loss is proposed,and then a joint loss function is constructed which combined with softmax loss to supervise the model to learn a deep feature space with stronger discriminative ability.In the learned feature space,homogeneous samples are compact with each other,and heterogeneous samples are far away from each other,which effectively improves the scene classification performance.(4)Since the available training samples are limited,it is difficult to learn a scene classification model with excellent performance.Therefore,proposing a few-shot remote sensing image scene classification method based on global-local relation network.On the basis of analyzing the spatial distribution characteristics and multi-scale characteristics of objects in remote sensing scene images,the method represents the input samples from two different scales,i.e.,global and key local,and has better adaptability when faced with diverse few-shot tasks.To effectively calculate the similarity of sample features,a relatively flexible and learnable deep relational network is used to calculate the similarity score.Finally,the category determination of the query sample is performed by combining the similarity scores of the two scales.