Identification Of Winter Wheat Based On Sentinel-2 Data In Heterogeneous Planting Conditions

It is of great importance to acquire winter wheat planting information for agricultural management,crop yield estimation and food security.Remote sensing technology is one of the effective means to extract winter wheat.Currently,the increasing satellite images available worldwide,such as Sentinel-2,GF-5,has created new possibilities for accurate maps of crop distribution.In this thesis,northern Anhui counties（Lingbi County,Sixian County）and central Anhui counties（Changfeng County,Dingyuan County）were selected as study areas.Based on the Sentinel-2,we applied decision tree,BP,SVM,RF and U-Net to extract winter wheat by considering the spectrum and phenology.The main contents and results of this thesis included:（1）The decision tree model was constructed to extract winter wheat based on NDVI variation trend in five key phenological periods,and the results were verified by statistical area and spatial distribution.Comparing the extracted area with the winter wheat sown area recorded in the statistical yearbook,the area percentage error of the northern Anhui counties is about 8%,and the central Anhui counties is about 20%;The ground samples were established by Planet images（spatial resolution 3m）,and the extraction results from Planet ground sample were used as reference data to verify the Sentinel-2 extraction results,the Kappa in northern Anhui counties was about 0.80,and the Kappa in central Anhui counties was between 0.60-0.65.The results showed decision tree algorithm can realize the winter wheat identification and achieve acceptable effect.（2）The study area belongs to the transition region between the north and the south,with frequent cloud coverage and limited available optical data.Therefore,the single period data has more advantages.Based on five key phenological periods,JM distance algorithm was used to determine the optimum period.Subsequently,three machine learning algorithms（BP,SVM,RF）were compared to screen the best classifier through the optimum period data.The result showed the heading stage（early April）was the optimum period for winter wheat identification,and the random forest performed best in the two study areas based on heading stage Sentinel-2 date,that is,the best classifier.（3）The 19 features（ten spectral bands+seven vegetation indices+water index+building index）generated from the heading stage Sentinel-2 data.The importance of all features was scored using the feature selection algorithm of random forest,and the optimum feature for winter wheat identification was screened.Then,all features were arranged in descending order according to the result of importance score,and the optimum feature subset were determined by using the sequential backward selection method.The results showed the optimum feature subset of northern Anhui counties was the first five features（NDVI,NDVI₆,band 11,band 2and band 8）with the highest importance score,and central Anhui counties was the first seven features（band 6,NDVI,GNDVI,band 2,band 11,EVI and band 4）.The random forest algorithm was used for classification based on the optimum feature subset,and the area percentage error was less than 5%and Kappa about 0.85 in northern Anhui counties,and the area percentage error was about 20%and Kappa ranging from 0.70 to 0.75 in central Anhui counties.Compared with the traditional band reflectance,the result of using the optimum feature subset data showed a higher accuracy of classification with a great advantage in data volume and processing time.（4）The identification of winter wheat was carried out by using U-Net in deep learning based on the optimum feature subset.The 20000 samples of winter wheat was derived from the Planet images（spatial resolution 3m）.Two thirds of the samples were used for U-Net training,and one third for model verification.The results showed the Kappa of the U-Net was about 0.87,and the random forest was about 0.85 in northern Anhui counties,the extraction effect of U-Net was better than random forest.In central Anhui counties,the accuracy of U-Net（Kappa was about 0.75）was also improved compared with random forest.On the whole,deep learning showed great potential in the fine recognition of winter wheat.