Research On Remote Sensing Image Object Classification Based On Graph Neural Network

As an important mean of obtaining information from remote sensing images,remote sensing classification plays an important role in many fields such as geographical condition monitoring,environmental protection and disaster assessment.In remote sensing images,they can be mainly divided into hyperspectral images and multispectral images with high spatial resolution.However,due to the imaging differences of remote sensing images,different types of images have different characteristics.In hyperspectral images,there are a lot of spectral information redundancy in high-dimensional spectral features,which is prone to dimension disaster.At the same time,there are some problems in hyperspectral images,such as few labels and unbalanced sample distribution,which pose great challenges to the classification of hyperspectral images.Compared with hyperspectral data,the spatial resolution of multispectral images is often higher and can obtain a wide range of data.However,multi-spectral images have a large number of spectral redundancy information in adjacent pixels,and the shape of different objects is different.It is difficult to capture the feature relationship between objects,which increases the difficulty of multi-spectral image classification.As a new technology,graph neural network is a new neural network model which combines graph data structure and convolutional neural network.It has the characteristics of local spatial perception,semantic features and spatial relationship constraints,which is beneficial to solve the problems of spectral information and spatial features in remote sensing object classification.In addition,graph structure is good at constructing spatial object dependencies while extracting rich spectral features in the data.Therefore,graph neural network provides a new solution for remote sensing image classification.In this thesis,according to the characteristics of hyperspectral and multispectral image data,the classification method of remote sensing image features is studied.By mining the characteristics of graph neural network and combining with geological knowledge,two graph neural networks are proposed to realize the accurate classification of hyperspectral and multispectral images,aiming at further improving the performance of remote sensing image feature classification.The specific contents of this thesis are as follows:(1)Aiming at the problems of dimension disaster and few sample labels in hyperspectral image data,this thesis proposes a multi-scale graph convolution network(MSGCN)based on spatial spectrum feature fusion to realize hyperspectral image feature classification.In this network,effective bands(dimension reduction)are selected from the spectral dimension through 1×1 convolution to obtain more abstract node features and edge features.In the process of graph convolution,this thesis uses the multi-scale convolution module to achieve the accurate classification of hyperspectral images.In addition,the multi-scale convolution module realizes the deep fusion of spatial spectrum features through multi-order proximity relations.In the experiment,three common hyperspectral data sets are selected to verify the effectiveness and robustness of the proposed method.(2)Multispectral imagery has noise and redundant information in spatial features,and is not good at capturing feature relationships at the object level.Inspired by U-Net and graph neural network,this paper proposes a U-object-oriented graph neural network(U-OGNN)to improve the performance of multi-spectral image classification.Firstly,considering the spectral and geometric characteristics,different uniform objects are extracted by super-pixel segmentation algorithm.Secondly,this kind of object is regarded as an entity node that is input into U-OGNN for context description;then,an adaptive composition strategy based on node depth features is designed to realize the deep fusion of spatial spectrum features by constructing a global relationship.Finally,a new U-shaped graph neural network structure is designed to capture abstract semantic features at different levels.Experiments show that compared with other models,the proposed method can achieve better performance on public high-resolution image dataset and Shandong land cover dataset.