Remote Sensing Identification Of Garlic In The Early Stage Based On Sentinel-2 Image

Timely and accurate identification of crop planting area is of great significance for food security and national policies,and is an important measure to assist agricultural power.Satellite remote sensing technology has become the most effective method for obtaining largescale crop distribution due to its advantages of low cost and high efficiency.Garlic is an important cash crop in China.It is of great significance to carry out research on key technologies of garlic mid early remote sensing identification for crop monitoring and agricultural insurance.However,the contradiction between the spatiotemporal differentiation of features in images of similar crops and the convergence of features in images of different crops in large regions remains a key bottleneck that restricts the accurate recognition of garlic crops through remote sensing.Specifically,the existing research area is small,making it difficult to generalize research results to large regional scales;(2)Realizing remote sensing recognition of garlic planting distribution in the early stages of garlic growth poses new challenges to key spectral and phenological features;(3)Significant phenological differences in large regions,diverse satellite imaging environments,and similar growth cycles of different crops such as garlic and winter wheat coexist.In addition,the fragmentation of garlic planting plots in China is high,and the problem of mixed pixel foreign objects being in the same spectrum is prominent,which easily leads to an increase in the uncertainty of crop image features,severely limiting the acquisition of reliable training sample points.Therefore,this article focuses on the need for early identification in garlic and the difficulty in distinguishing between garlic and winter wheat.Using Sentinel-2 images with a 10 meter spatial resolution as the data source,Henan and Shandong provinces as the research areas,and relying on the Google Earth Engine(GEE)cloud platform,we conduct research on key technologies for precise remote sensing recognition of early and middle stage garlic.Firstly,analyze the feasibility of early identification in garlic;Then,construct the garlic image enhancement index and the winter wheat image enhancement index,and superimpose them to generate an image enhancement composite image;Finally,coupling the image enhancement composite image with the GVI composite image to improve the classification accuracy of garlic.This article conducted a distribution mapping of garlic four months before its harvest period,with an overall classification accuracy of 95.19% and a Kappa coefficient of 0.89.The producer accuracy of garlic was 97.50%,and the user accuracy was 94.92%.This article addresses the key challenges faced by remote sensing recognition of garlic on a large scale in geographic space,achieving mid to early identification of garlic,and achieving the following innovative results:(1)Based on spectral characteristics and activation function,Garlic Image Enhancement Index(GIEI)and Wheat Image Enhancement Index(WIEI)were constructed.In this study,Sentinel-2 image is used to analyze the spectral characteristics between garlic and other ground objects.It is found that the spectral differences between garlic and winter wheat are mainly in red,green,blue and NIR bands,and the reflectivity of garlic in red,green and blue bands is higher than that of winter wheat.The reflectivity of garlic in near infrared band is less than that of winter wheat.The spectral difference between garlic and bare land and deciduous forest is from near infrared band to the SWIR1 band.Garlic is on the rise,while bare land and deciduous forest are on the decline.Therefore,in this paper,the activation function and four operations are used to normalize the feature factors,and the feature values are compressed to(-1,1)to generate the garlic image enhancement index and the winter wheat image enhancement index respectively,and the two index images are superimposed to generate the image enhancement composite image.The overall classification accuracy of garlic calculated by Iso Data classification method is 91.03%,and the Kappa coefficient is 0.82,which is 4.23% and 0.1 higher than the original image respectively.(2)Built a Green Vegetation Index(GVI)to accurately identify potential distribution areas of garlic.In order to reduce the misclassification of garlic with wasteland and construction land,this paper constructs a green and Green Vegetation Identification(GVI)based on Sentinel-2image,and uses GVI time series to determine three effective time windows.In the first highvalue time window of GVI(from July 10,2020 to September 1,2020),a composite image of the maximum value of GVI is generated,which is defined as GVI＿max＿8.Generate a GVI median composite image within the low-value time window of GVI(from October 1,2020 to October 30,2020),which is defined as GVI＿med;In the second high-value time window of GVI(November 20,2020 to January 20,2021),the composite image of GVI maximum is generated,which is defined as GVI＿max＿12.Based on the synthetic images of three windows,the differences between garlic and other ground objects are analyzed,and the final GVI synthetic images generated by GVI＿max＿12,the sum of GVI＿max＿8 and GVI＿max＿12,and the differences between GVI＿max＿12 and GVI＿med.The overall classification accuracy,Kappa coefficient and F1 score of the potential distribution area of garlic obtained by calculating GVI composite images with Iso Data classification method are 93.13%,0.86 and0.93 respectively.(3)Coupling enhanced index synthetic image and GVI synthetic image,remote sensing mapping of garlic distribution was achieved 4 months before harvest.In order to improve the classification accuracy of garlic,the enhanced index composite image and GVI composite image are coupled in this paper.The overall classification accuracy of garlic is 91.03% and Kappa coefficient is 0.82 by using Iso Data classification method.Comparing the effects of decision tree,random forest and support vector machine on garlic identification,it is found that the decision tree has the highest classification accuracy,with the overall classification accuracy of 95.19% and Kappa coefficient of 0.89,among which the producer accuracy of garlic is 97.50% and the user accuracy is 94.92%.