Classification Of Typical Wetlands In The Middle Reaches Of Tumen River Based On Sentinel Image

As one of the three major ecosystems of the earth,the wetland ecosystem has functions such as climate regulation,soil and water conservation,and water purification.At the same time,it is also a habitat for various rare animals and plants,which is of great significance to human production and life,and is known It is the "Kidney of the Earth".Despite the importance of wetlands,the statistics of wetland dynamics on a global scale show a continuous downward trend.Therefore,scientific research on wetland monitoring is imperative.Remote sensing technology is an important tool for monitoring wetland patch distribution,vegetation species differentiation and growth status in an area.Wetland resource survey and land cover dynamic monitoring based on remote sensing and GIS technology have become indispensable methods for wetland scientific research and management.This paper uses the 2019 Sentinel-2B multispectral imagery and Sentinel-1A radar imagery to integrate the advantages of multi-source remote sensing images to establish a local classification system,extract multiple feature variables,and use object-oriented classification methods.Accurate extraction and comparison of wetland information in the study area fills the gaps in the classification of multi-source remote sensing images in this area,and provides important data support for ecological protection and wetland research in the Tumen River Basin.The main research content and results of this paper are as follows:(1)The multi-spectral images obtained by Sentinel-2B optical satellite and the radar images obtained by Sentinel-1A radar satellite are used as the basic data for wetland land cover classification research.The comprehensive spectral feature information and radar features with good vegetation penetration and not affected by weather are utilized to realize the complementation of remote sensing information.Firstly,based on field investigation,a classification system suitable for the study area was established,including land cover types such as woodland,forest swamp,herbaceous swamp,dry land,reservoir,construction land,etc.Secondly,object segmentation parameters were determined through experiments and calculation of the optimal segmentation scale evaluation tool.Combined with the selection of the optimal feature variables,spectral features,geometric features,texture features,radar features and terrain features were selected.Finally,the random forest classification method is used to identify and extract the information with the object as the processing unit,which provides a scientific basis for the wetland classification research of Tumen River Basin and the advantages of integrating multi-source remote sensing data.(2)According to the statistics of land cover classification data,the overall classification accuracy is 83% and the Kappa coefficient is 0.78,which indicates that the land classification accuracy in the study area is high and the classification results are highly correlated.The main land cover types in the study area were forest land,forest swamp and dry land,which accounted for about 94% of the total area of the study area.The producer accuracy was 90%,88% and 87%,respectively.The producer accuracy of construction land and reservoir reached 92%.The overall results showed that the random forest classification based on object orientation could well distinguish the forest swamp from the herbaceous swamp,and the continuity and integrity of ground features were relatively high.At the same time,according to this study,the addition of polarization characteristics and backscattering coefficient of radar data can improve the misclassification and mixing phenomenon of forest swamp and other ground object types,and can reflect the ground object information more objectively and truly.(3)Compared with other classification methods,the overall classification accuracy and Kappa coefficient of the object-oriented random forest classification method based on multi-source remote sensing images are higher,and the classification accuracy of ground object information is better.Compared with the object-oriented classification based on Sentinel-2B images,it is found that in the absence of radar image features,the classification accuracy of some land cover types,such as forest marshes and herbaceous marshes,will be reduced by relying only on multi-spectral image classification.Compared with the random forest classification based on Sentinel-2B image pixel,it is found that the "pepper and salt" phenomenon is more serious because of the method based on pixel as classification unit.The objects generated by the multiscale segmentation algorithm are to gather the pixels of the same kind together,which not only maintains the basic features of the pixels,but also generates new spatial information,and significantly improves the classification accuracy.Compared with the object-oriented random forest classification based on Landsat-8 images,it is found that the classification effect of this method is the worst,indicating that the spatial resolution and radiative resolution of remote sensing data sources play a crucial role in the classification accuracy of the study area.In some cases,it may affect the classification accuracy more than the classification method.