| Winter wheat is one of the main food crops in the world,and accurate knowledge of its spatial distribution is of great significance for ensuring national food security.How to quickly and accurately extract the spatial distribution of winter wheat is a common concern of researchers.With the development of remote sensing technology,medium-resolution remote sensing images have become the main data source to obtain the spatial distribution of winter wheat in a large range.However,because the characteristics of winter wheat planting areas are not obvious on medium-resolution images,extracting its high-quality characteristics is the key to extract the accurate spatial distribution of winter wheat.In this paper,Sentinel-2A image is selected as the data source.Firstly,the channel downscaling model is constructed to solve the problem of data downscaling in order to solve the problem of red edge and short-wave infrared band and visible light band.For the problem of obtaining high-quality features,a pixel-by-pixel segmentation model is constructed,focusing on solving the problem of extraction accuracy.Mainly completed the following work:1.Data set making.Sixty-two Sentinel-2A remote sensing image data of winter wheat in 2022 were selected,and the remote sensing image processing software was used as the image processing tool.After preprocessing steps such as radiometric calibration,atmospheric correction,orthorectification and geographical registration,combined with previous research experience and visual interpretation results,the winter wheat area was manually delineated,and more than 1,000 sets of sample data were produced.2.Analyze Sentinel-2A data characteristics.With PIE(Pixel Information Expert-Basic)software as the analysis tool,the characteristics of band such as minimum value,maximum value,mean value,range,standard deviation,histogram,covariance matrix and correlation coefficient matrix,as well as six vegetation indices such as NDVI,NDWI,RVI,EVI,NDRE1 and NBR were analyzed by statistical analysis method.The features such as B2,B3,B4,B5,B6,B7,B8,B11,EVI,NDVI and NDRE1 are selected as the basic features,which are indicative for identifying winter wheat.3.Build channel downscaling model.Aiming at the problem that the spatial resolution of the red edge and short-wave infrared band in the Sentinel-2A image is inconsistent with that of the visible light band,a channel downscale model REDS(Red Edge Down Scale)is constructed based on deep learning.By injecting the spatial structure information extracted from the highresolution channel into the low-resolution channel,the model achieves the purpose of obtaining high-resolution red edge channel and SWIR channel data,and finally achieves the goal of reducing the spatial resolution of red edge and short-wave infrared band from 20 m to 10 m.The experimental results show that the structural similarity index,peak signal-to-noise ratio and spectral mapping angle of REDS method are 0.92,29.05 and 3.27,respectively,which lays the foundation for the next work.4.Build a pixel-by-pixel segmentation model.Aiming at the problem of obtaining highquality features,a pixel-by-pixel segmentation model,REVINet(Red Edge and Vegetation Index Feature Network),is constructed based on the convolutional neural network.The model takes B2,B3,B4,B5,B6,B7,B8,B11 with 10 m resolution as basic input data,and the extracted combinations of enhanced vegetation index,normalized vegetation index and normalized difference red edge index,so as to obtain accurate spatial distribution of winter wheat.The experimental results show that the recall,precision,accuracy and F1 score of REVINet model reach 92.47%,93.99%,93.34% and 93.22% respectively,and the area accuracy of winter wheat extracted in the study area is 94.23%,which shows that this method has obvious advantages in extracting the spatial distribution of winter wheat. |
PDF Full Text Request