Accurate Recognition Algorithm Of Winter Wheat Based On Sentinel-2 And Landsat8 Image

Winter wheat is one of the main food crops in northern China.Its growth,pests and diseases,and planting area affect the national food security to a large extent.It is of great significance to grasp the planting situation of winter wheat timely and accurately,pay close attention to the trend of winter wheat planting area around the country,ensure the domestic food supply,keep the safety bottom line of 1.4 billion people in China,formulate food policy as soon as possible,prevent crop disasters,and provide agricultural subsidies.Remote sensing technology has become a very important technical means in the development of fine agriculture and intelligent agriculture because of its convenient and rapid acquisition.However,the accuracy of crop extraction by remote sensing technology is usually affected by many factors.However,the accuracy of crop extraction by remote sensing technology is usually affected by many factors.Most researchers only consider improving a certain factor to improve the extraction accuracy of winter wheat,and have not systematically studied the improvement of data quality,sample selection method and classification algorithm performance.To solve the above problems,the main contents and conclusions were as follows:(1)Aiming at the problem that the quality of image data affects the extraction accuracy,the advantages of Sentinel-2 data and Landsat8 data were fully exploited.Three kind of characteristics of Sentinel-2 data such as reflectivity,vegetation index and texture were introduced to fuse landsat8 data.The fusion effect of Gram-Schmidt(GS),NNdiffuse Pan Sharpening(NN)and PC Spectral Sharpening(PC)were analyzed by combining qualitative evaluation with quantitative evaluation(mean,standard deviation,information entropy).The results showed that the information entropy of the image(NN-TEXTURE1)using NN fusion method under the first component of texture feature was 14.52,which was 3.01,1.62 and0.68 higher than that of Sentinel-2,Landsat8(30 m)and Landsat8(15 m),respectively.All indicated that NN-TEXTURE1 can fully combine the advantages of the two images data to improve the data quality.(2)The sample selection method affects the extraction accuracy of winter wheat,two sample selection methods were based on regional samples and endmember samples.Experiments were carried out on six traditional supervised learning classification algorithms,namely Mahalanobis Distance Classification(MHDC),Maximum Likelihood Classification(MLC),Minimum Distance Classification(MMDC),Random Forest Classification(RFC),Spectral Angle Mapper(SAM)and Support Vector Machine(SVM),and three groups of training samples.The classification accuracy was evaluated by Kappa,overall accuracy and extraction area.The results showed that the area of winter wheat obtained by endmember training samples was 47000～55000hm~2,and the area of winter wheat obtained by regional training samples was 60000～69000hm~2.Compared with the official statistical data of58333hm~2,the classification area of endmember samples was smaller,and the regional samples was larger.Under the regional samples,the MLC Kappa was 0.91,with the highest classification accuracy,and the extraction area was 60747hm~2,which was closest to the actual planting area.(3)To improve the extraction precision of winter wheat in small plot,Our study proposed to add the Atrous Spatial Pyramid Pooling(ASPP)to the encoding and decoding part of the U-net network and replace all ordinary convolutions with the Dilated Convolution(DC)to form ASPP-U-net.The experimental results showed that the Kappa of ASPP-U-net was 0.93,which was 0.01 and 0.02 higher than that of U-net and MLC,respectively.Through mapping comparison and analysis with MLC,it was found that the winter wheat plots identified by ASPP-U-net were more complete and clearer than those identified by MLC,which could realize the accurate identification of winter wheat at plot scale.