Collaborative Classification Of Land Cover Based On Polarimetric SAR And Optical Imagery

Land cover classification can intuitively reflect the current situation of natural and human development in a region,and is an important subject of current remote sensing image research.With the advancement of remote sensing acquisition technology,the acquisition of multi-source and multi-temporal remote sensing image data has become more and more convenient.The application of traditional optical remote sensing is affected to a certain extent due to the constraints of external factors such as weather and light.Synthetic Aperture Radar(SAR)belongs to the category of active remote sensing and can work all day and all day.As one of the current hotspots in the study of remote sensing images,the use of SAR images for classification research is highly valued.This paper takes the vicinity of Nanjing as the research area,and uses the 6 phases of Sentinel-1 and Sentinel-2 images of ESA,based on the polarization characteristics,backscattering coefficients,interference characteristics and spectral characteristics of optical images,and normalized differential vegetation based on SAR polarization characteristics,backscatter coefficients,interference characteristics and optical image spectral characteristics.Index and texture features,using random forest algorithm to optimize the selection of many feature variables,and finally put the selected feature variable combination into the random forest classifier for classification and accuracy evaluation,to study the feasibility of the application of multi-source remote sensing image features to land cover classification Sex.The main research contents and conclusions of this paper are as follows:(1)Analysis of time series characteristics based on SAR images.For six types of ground objects,the backscatter coefficients of single-phase SAR images are compared with time series backscatter coefficients and coherence coefficients for theoretical analysis,and the discriminant analysis in mathematical statistics is used to verify the accuracy of the theoretical analysis.In the single-phase SAR image data,the backscattering coefficients of vegetation-like ground objects such as cultivated land,garden land,and wetland are similar,but the difference is small.The single-phase backscattering coefficient alone cannot be effectively distinguished.Add the time series SAR image The classification of the backscatter coefficient and coherence coefficient can significantly improve the accuracy of SAR image classification.(2)Optimized screening based on the characteristics of multi-source remote sensing images.Extract the spectral information,time series NDVI and texture feature information of the ground objects in the optical image,the covariance matrix in the SAR image and the polarization characteristics after polarization decomposition,and conduct the theoretical analysis of the separability of various ground objects,and adopt The random forest algorithm sorts all the features by feature importance,reduces the dimensionality of the high-dimensional feature variable space according to the feature importance score,and selects the top-ranked feature variables to participate in the final image classification.SAR image time series backscatter coefficient,coherence coefficient and polarization characteristics,optical image spectral characteristics,time series NDVI and texture characteristics,participating in the collaborative classification of SAR and optical images,can reach an overall accuracy of 88.4%,compared with single-phase SAR The accuracy of image classification is increased by 9.1%,which strongly proves that the theoretical analysis of SAR and optical image information complementation is more reliable.(3)Research on classification based on multi-source time series remote sensing images.The random forest algorithm is used to classify the images in the study area by four methods:single-phase SAR image classification,time series SAR image classification,optical feature variables participating in the classification and feature variable screening and optimization,and finally combining reference images and field data to classify the four Method for accuracy evaluation.After the random forest algorithm screens various features of SAR image and optical image,the texture feature of optical image is outstanding when participating in the classification of multi-source remote sensing image.The SAR polarization image feature~-?/A after the band calculation is participating in the classification.The importance of time is also in the forefront,consistent with the theoretical analysis,indicating that the theory of random forest algorithm for screening and optimizing multi-source image features is feasible and can effectively improve the performance of the classifier.After using random forest to optimize the input variables,the highest overall accuracy is 92.5%,and the classification accuracy is 13.2% higher than that of single-phase SAR image classification,which reflects the advantages of combining optical image and SAR image classification;at the same time,it is compared with the classification of all feature variables.Method,the results after feature screening are more in line with the actual feature distribution,and the classification details are more prominent.