Scene Classification And Extraction Of High-resolution Remote Sensing Image Based On Lie Group

With the rapid development of remote sensing technology and various sensor platforms,high-resolution remote sensing images(HRRSI)are easier to obtain than before.How to efficiently utilize the increasingly rich feature information to accurately classify and extract HRRSI scenes is one of the key issues in remote sensing image interpretation.Although from pixel-based method to object-based scene classification and extraction method,the classification accuracy of ground objects has been greatly improved.However,the traditional method model can only achieve the classification and extraction of ground objects,such as aircraft,trees,and buildings,while it is difficult to achieve the classification and extraction of high-level semantic levels such as urban villages(slums)and industrial parks.How to break through the problems existing in the existing methods and models is a hot,difficult,and key problem in the field of remote sensing.At present,the combination of a deep learning model for remote sensing image scene classification and extraction is a hot research issue.However,this kind of model has the following shortcomings:(1)The existing convolutional neural network(CNN)models use weight sharing,which leads to the loss of local feature information.Two HRRSI with different global features may belong to the same scene because they contain some common salient objects.In this case,relying solely on global features may lead to misclassification.(2)Although some models consider both local and global features as the final representation,they do not consider the relationship between different regions and the relationship between different features in the image.In addition to the attributes of features and objects themselves,the internal relationships between them are also very important for classification and extraction.(3)The existing CNN model is usually regarded as a ”black box”,and the decision of the model is not transparent and interpretable,and the structure of the model is becoming more and more complex with more parameters.In addition,most of the existing models are easy to lose the spatial location information,and can not extract more effective and distinguishing features for the complex spatial distribution and geometric structure of the scene.The model based on feature learning mainly represents the HRRSI by extracting various features,which is one of the effective methods to solve the problem of classification and extraction.However,it is more challenging for scene classification and extraction in HRRSI:(1)Traditional feature learning methods only utilize a single feature,and the representation ability is insufficient,while multi-feature learning simply accumulates features,and the correlation and difference between features are not fully expressed and learned.(2)Compared with natural images,the scene objects in HRRSI are more complex,with large intra-class variance and small inter-class variance.(3)In HRRSI,the scale of the scene target changes greatly,which causes the multi-scale problem and makes the scene target localization and recognition more difficult.Given the problems in the above method model,this paper introduces the Lie Group Machine Learning method,reuses the features of HRRSI,and proposes a scene classification and extraction method based on Lie Group Machine learning.Taking HRRSI as an example,it carries out quantitative and qualitative experimental analysis and demonstration.The main research contents and innovative work are as follows:(1)The characteristics of HRRSI and their scenes are deeply analyzed,the existing research methods,trends,and problems of scene classification and extraction are summarized,and the basic theories and methods related to this study,such as Lie Group Machine Learning,compressed sensing,and graph cutting algorithm,are introduced.(2)In the aspect of distance classification and mean algorithm,aiming at the problem that the existing algorithms based on Euclidean distance calculation can not accurately express the non-Euclidean space data samples,this paper introduces the intrinsic mean classification algorithm on Lie Group manifold space for scene classification.To address the problem that the existing kernel functions are based on vector space samples and can not be directly applied to matrix space samples,a Lie Group kernel function satisfying both matrix samples and vector samples is derived in this paper.In this paper,a scene classification algorithm is proposed,which combines the intrinsic mean of Lie Group and kernel function of Lie Group.The Euclidean distance is replaced by the space distance of the Lie Group manifold,which improves the accuracy of scene classification simply and effectively.After the optimization and improvement of the algorithm,a lightweight intrinsic mean scene classification algorithm based on Lie Group kernel function is obtained,which reduces the time complexity of the algorithm and further improves the accuracy of scene classification.(3)In the aspect of model mechanism interpretation,the existing convolutional neural network(CNN)is regarded as a ”black box” and its decision-making is not transparent,so the correlation representation between image features and scene semantics is established in this paper;in the perspective of the lack of global,local regional and spatial correlation information in CNN,the location coding method is used to enhance the spatial information,and the low-level,middle-level and high-level features are used fusion representation of remote sensing scene.In this paper,a scene classification method based on the convolution network multidilation pooling of Lie Group is proposed.The model improves the interpretability of the model from the perspective of Lie Group feature learning.The features of different regions(global and local),different layers(low-level,middle-level,and high-level),and their correlation information are encoded into the Lie Group region covariance feature matrix,which has the advantages of fewer parameters,low dimension,anti-noise,and good computing performance.(4)In the aspect of the universality and robustness of the model algorithm,the CNN model has the problems of complex model structure,more parameters,and high computational complexity due to the large intra-class variance and small inter-class variance in the scene.This paper based on the improved Fisher dictionary feature coding algorithm,better balances the problems of accuracy,parameter size,and computational performance,and the algorithm can deal with different manifolds Structural data,with good robustness and universality.Aiming at the limitation of the traditional Grabcut algorithm,which constructs a Gaussian mixture model based on pixel brightness value and can not be directly applied to the sample of Lie Group matrix,an improved Grabcut extraction algorithm is proposed.The algorithm utilizes trace based on the regional covariance feature matrix of Lie Group to construct a Gaussian mixture model,which enhances the universality and robustness of the algorithm.At the same time,to address the problem of iterative calculation in the algorithm,the algorithm introduces the divergence of KL and the information entropy of the image,taking into account the computational efficiency and extraction effect.In this study,the construction and learning of regional covariance feature matrix based on Lie Group,the interpretability of model mechanism,the improvement of accuracy and robustness of algorithm model are carried out,the bridge between scene semantics and image features are established,the application of remote sensing image scene classification and extraction,in reality,is accelerated,and the important scientific reference value is provided for urban planning and other fields.