Research Of Land Cover Classification Based On Graph Convolutional Neural Network

The extraction of land cover information is very important for environmental monitoring,resource investigation and other research,and it plays an important role in the formulation of national macro-control and related policies.With the rapid development of satellite sensing technology and aerial photography technology,there are more and more remote sensing image acquisition methods,and the spatial resolution is getting higher and higher.It provides more detailed spectrum and texture information of ground objects.The phenomenon of different spectra and the same spectra of foreign objects is more serious,which brings greater challenges to land cover classification.Although there have been a large number of pixel-based classification methods used in land cover classification,it is difficult for these methods to achieve high accuracy and efficiency in the classification of high-resolution remote sensing images.Different from the traditional pixel level,the object-oriented image analysis method uses the object as the basic research unit,which provides a new research idea for the extraction of high-resolution remote sensing information,and achieves the purpose of reducing salt and pepper noise and improving processing efficiency.However,the spatial relationship characteristics of image objects are difficult to be effectively analyzed and expressed in object-oriented classification methods,which restricts the accuracy of land cover classification and the depth of analysis and understanding to a certain extent.At the same time,the update iteration of deep learning technology provides a powerful means for the automatic and efficient interpretation of remote sensing image information,bringing unprecedented opportunities.In response to the above problems,the main content and contributions of this article are as follows:(1)In order to solve the defect that the conventional segmentation methods cannot make full use of the edge information of the features in the segmentation process,this paper uses the RCF(Richer Convolution Features)model to learn the ability of the edge features of the features,A multi-scale segmentation method based on deep supervision is constructed,which makes full use of the edge information of features as supervision and constraints in the segmentation process,which makes up for the shortcomings of traditional segmentation.Compared with other first-class segmentation methods,the segmentation results have more compact edge details and more consistent with actual features.This fully demonstrates the effectiveness of this method.(2)Since the characteristics of the image determine the accuracy of land cover classification to a certain extent,this paper uses the deep feature extraction capabilities of the convolutional neural network(CNN)based on the image object to extract the deep features of the object at different scales.The image objects are regarded as nodes,the adjacency relationship between the objects is calculated,and the deep features of the objects are combined to construct a graph data set for subsequent classification.(3)In order to analyze and express the spatial relationship of image objects,this paper makes full use of the powerful classification ability of graph convolutional neural network(GCN)and the learning ability of topological relationship between nodes,and proposes an object-obsed graph convolutional neural network(OGCN).The method of extracting land cover information from remote sensing with high spatial resolution.Use the constructed graph data set to train GCN and predict the result of land cover classification through the trained model.In summary,this article explores the impact of object feature extraction scale on classification results through experiments,and demonstrates the importance of segmentation methods and classification methods.Through the data set s1 and data set s2,we compared our proposed method with the first-class segmentation and classification methods.The results show that our method has a more complete contour of the ground features in image segmentation,and the final land cover classification accuracy is different.It is 0.834 and 0.903,which is better than other methods.