| Human activities are frequent in the coastal areas of the Guangdong-Hong Kong-Macao Greater Bay Area(GBA),and coastal wetlands are important ecological resources in the region.It is of great significance to use remote sensing technology to classify and monitor their resource information with high precision.At present,in the research on remote sensing classification of coastal wetlands in the GBA,there are few studies on new machine learning methods,medium and high-resolution and multi-source remote sensing data,and regional classification feature variables,and there is not yet a complete set of high-precision coastal wetland information extraction in remote sensing program.This paper takes the typical coastal wetlands in the GBA as the research object,combined with the geographical environment characteristics of cloudy and rainy climate,rich moisture information,and complex surface coverage,from the perspective of remote sensing data sources,classification feature variables,and remote sensing classification methods.High-precision extraction of typical coastal wetland information in the GBA.The main contents and conclusions of this paper are as follows:(1)Differences in classification accuracy of optical and fusion remote sensing data in information extraction of coastal wetlands in the GBA The climate in the study area is cloudy and rainy,and the moisture information is abundant.Single optical remote sensing data cannot meet the requirements of high-precision information extraction of coastal wetlands.Based on the random forest algorithm,this paper selects three commonly used and freely available medium-high spatial resolution remote sensing data—Landsat 8(30 m),Sentinel-2(10/20 m),Sentinel-1(10 m),combined with two commonly used optical and Synthetic Aperture Radar(SAR)data fusion methods—principal component(PC)and wavelet transform(WT)fusion,construct and obtain six fused and non-fused datasets,and explore the differences between multi-source remote sensing data in the information extraction of coastal wetlands in the GBA.The results show that the overall classification accuracy of Sentinel-2 optical data source is better than that of Landsat 8 optical data source,and the fusion data of Sentinel-2 optical and Sentinel-1 SAR wavelet transform can achieve the highest classification accuracy(the overall accuracy is 91.37%,and the kappa coefficient is 0.9003).Compared with Landsat 8 optical remote sensing data,Sentinel-2 optical remote sensing data with higher spatial resolution and richer spectral information have higher information extraction accuracy for coastal wetland types with complex surface coverage and similar spectral information,such as shallow waters,estuaries Waters,coastal tidal flats,aquaculture ponds,reservoir ponds and mangrove wetlands.(2)The contribution of different data fusion methods to the classification accuracy of coastal wetlands in the GBA.The choice of optical and SAR image fusion method directly affects the quality of the fusion image and the final classification accuracy.Based on two fusion methods,principal component and wavelet transform,this study conducts fusion experiments on Landsat 8,Sentinel-2 optical and Sentinel-1 SAR images.The results show that the wavelet transform fusion of optical and SAR images can improve the information extraction accuracy of coastal wetlands in the GBA,while the principal component transform fusion method has little improvement in the information extraction accuracy of coastal wetlands(p < 0.05).It is worth noting that in the principal component transformation fusion,the addition of SAR data helps to distinguish water body and non-water body information,and can alleviate the misclassification of coastal tidal flats and turbid water bodies,mangrove wetlands and nonwetland farmland.(3)Important characteristic variables for high-precision extraction of coastal wetland information in the GBA.Feature variables are an important basis for machine learning methods to achieve land object classification.The random forest algorithm used in this study can evaluate the importance of characteristic variables in classification.The results show that:terrain(elevation,aspect,slope),water index(MNDWI,NDWI,TCW),vegetation index(NDVI,RVI,DVI,CIgreen)are important features for information extraction of coastal wetlands in the GBA,of which the importance of water body index features is higher than that of vegetation index features.Optical image visible light and near-infrared band reflectance,RVI,TCW,altitude,Sentinel-1 VV polarized image contrast,NDVIre2,Sentinel-1 VV polarized backscatter coefficient are the most important specific characteristic variables among various characteristics.Sentinel-1 VV co-polarization backscattering coefficient intensity is higher than Sentinel-1 VH cross-polarization backscattering coefficient intensity,which is more suitable for the extraction of coastal wetland information in GBA.(4)The performance of different remote sensing classification methods in information extraction of coastal wetlands in the GBA.This paper uses the data set with the highest classification accuracy and its 26 characteristic variables as the data source,based on three common machine learning methods,to explore the performance differences of different classification methods in the information extraction of coastal wetlands in the GBA.The results show that the overall accuracy of the random forest algorithm is significantly higher than that of the support vector machine method and the maximum likelihood method(p < 0.05).The overall classification accuracy of the random forest algorithm is the highest(the overall accuracy is 91.94%,the kappa coefficient is 0.9069),the support vector machine method is second and more stable(the overall accuracy is 88.55%,the kappa coefficient is 0.8676),and the maximum likelihood method has the lowest overall accuracy(The overall accuracy is85.16%,and the kappa coefficient is 0.8287).In general,in the coastal area of the GBA with a cloudy and rainy climate,rich moisture information and complex surface coverage,based on the random forest algorithm,Sentinel-2optics and Sentinel-1 VV co-polar radar wavelet transform are fused as remote sensing.The data source,combined with the features of terrain,water body index,vegetation index,texture,radar backscatter coefficient and Sentinel-2 red edge index,can achieve high-precision extraction of typical coastal wetland information in the region.In the extraction of specific coastal wetland types,when natural costal wetland types were extracted with single Sentinel-2data and constructed costal wetland types were extracted with fusion data,the data set cost was lower and the classification accuracy was higher.This study can provide a reference for remote sensing classification data sources,method and characteristic variable selection for the coastal wetland resources survey in the GBA. |
PDF Full Text Request