Study On Crop Classification Based On Sentinel Data At A County Scale

Agricultural production is the basic guarantee for human survival and the leading factor of economic development.Crop planting information is the main basis for adjusting agricultural production activities and an important premise and basis for grain yield prediction and disaster assessment.Timely mastering the sowing area and spatial distribution information of crops is of great significance to ensure national food security,complete the carbon neutralization goal and ecological environment construction.In view of the low accuracy of extracting crop information in areas with broken plots and complex planting structure,thesis takes nangong city as the research object,uses Sentinel-1 SAR and Sentinel-2 multispectral data to construct a variety of backscattering coefficient and vegetation index time series,and extracts the key time phases of crop growth period combined with phenological information,A variety of classification schemes were combined to refine the extraction of typical crops in the study area.The effects of different feature combinations on the accuracy of crop classification were discussed in order to provide a method reference for the study of fine classification of crops at the county level.The main research contents and conclusions are as follows:（1）Based on Sentinel-2 optical images,five time series sets including Normalized Difference Vegetation Index（NDVI）,Red Edge Normalized Difference Vegetation Index(NDVI₇₀₅),Enhanced Vegetation Index（EVI）and two combination schemes(EVI+NDVI,EVI+NDVI+NDVI₇₀₅)were constructed.The harmonic characteristic parameters in each time series curve are solved by Fourier series,and the crop planting information in the study area is extracted by random forest classifier.The results show that the overall classification accuracy of EVI+NDVI+NDVI₇₀₅time series set based on exponential features and harmonic features is the highest,with a value of 86.38%.Compared with EVI+NDVI time series set,the overall classification accuracy is improved by 2.04%,indicating that the red edge index NDVI₇₀₅can effectively improve the identification ability of crops.Compared with the five time series sets based on exponential features,the overall classification accuracy of the five time series sets based on exponential features and harmonic features is improved,indicating that the addition of harmonic features can effectively improve the classification accuracy.（2）The time series curves of VV and VH backscattering coefficients were constructed using Sentinel-1 SAR data,and the crop planting information was extracted by twdtw and random forest classifier.The results show that the extraction accuracy of twdtw time series classification scheme is higher than that of random forest time series classification scheme.The overall classification accuracy of VV time series set,VH time series set and VV+VH time series set is improved by 3.86%,3.13%and 5.36%respectively.（3）Using the data of exponential feature,harmonic feature and polarization feature,the classification scheme is constructed by feature level fusion.The results show that the overall classification accuracy of the fusion scheme based on dual polarization feature,exponential feature and harmonic feature is the highest,reaching89.27%,and the kappa coefficient is 0.88.Under the best classification scheme,the planting areas of winter wheat summer corn,cotton,millet and other crops in nangong city in 2019 are 21580.77,16968.16,3000.81 and 403.36 hm²respectively.