Study On Extraction Of Fractional Vegetation Cover Of Typical Crop From UAV And Sentinel-2 Satellite Remote Sensing And Validation Of The Sentinel Simplied Level 2 Product Prototype Processor(SL2P)

Fractional vegetation cover(FVC),defined as the percentage of the vertical projection area of vegetation on the ground to the total statistical area,is not only an important indicator for describing the distribution of surface vegetation,but also an important parameter in ecosystem,soil erosion and climate change models.Although the accuracy of FVC obtained by ground measurement method is high,it is difficult to obtain FVC at regional scale.Unmanned aerial vehicle(UAV)remote sensing can meet the requirements of sampling accuracy and can also greatly save the working time,human and material resources,which is suitable for field sampling with large workload and wide range of collection.UAV can provide data with high accuracy and representative large quadrat for remote sensing estimation of FVC,thus improves the accuracy of FVC estimation model.Validation of remote sensing is to check the accuracy of products and algorithm to ensure the quality of satellite products and algorithms.SL2P is an estimation method of vegetation physicochemical parameters(FVC,leaf area index,etc.)based on sentinel-2 satellite data.The evaluation of this algorithm can provide a basis for its subsequent application.In this study,UAV remote sensing images with and without shadow influence were obtained by field experiments.FVC was extracted by bimodal histogram threshold method and OTSU threshold method respectively.For UAV remote sensing images with shadow influence,a double threshold method was proposed to eliminate shadow influence and achieved accurate extraction of FVC.Then,the validation of SL2P algorithm was verified by the extracted FVC data.Finally,the random forest algorithm and gaussian process regression algorithm were used to develop FVC estimation models.The evaluation indices selected in this study were extraction error(E_F),determination coefficient(R~2),root mean square error(RMSE),and estimation bias(Bias).The main conclusions of this study are as follows:(1)Firstly,based on the visible images of three crop types,i.e.,soybean,maize and paddy rice,EXG,VDVI,NGRDI,NGBDI,RGRI and EXG-EXR were selected.FVC was extracted by bimodal histogram threshold method and OTSU threshold method.The results showed that the bimodal histogram threshold method could accurately extract FVC from UAV images without shadow influence(average E_F=7.37%;RMSE=5.83%),OTSU threshold method was less accurate than bimodal histogram threshold method(average E_F=23.16%;RMSE=16.67%).Secondly,the above methods were used to extract FVC from UAV images with shadows.The results showed that the bimodal histogram threshold method was not suitable for the UAV images with shadow influence.The accuracy of OTSU threshold method was low(E_F=55.46%),which could not accurately extract FVC.(2)The UAV visible images under the influence of shadow has the following characteristics:for vegetation pixel,G>R and G>B,for light soil pixel,R>G>B,for shadow soil pixel and vegetation rendering soil pixels,G>R and G>B.Therefore,it will cause misclassification by simply using G>R and G>B to distinguish vegetation and soil.In this study,Lab color space and HSV color space were selected studying the method to accurately extract FVC from UAV images with the influence of shadows.The results showed that the a-channel fitting curve of Lab color space presented a bimodal distribution in high FVC region.Combined with the threshold method,the method could effectively remove the influence of shadow,and the extraction accuracy of FVC was high(E_F=6.58%,RMSE=5.21%),but in the middle and low FVC region,the a-channel fitting curve didn’t present a bimodal distribution,the FVC couldn’t be extracted.In HSV color space,H components of vegetation,shadow and soil were significantly different,and the influence of shadow was effectively eliminated by setting two fixed thresholds(0.253 and 0.3601).The two fixed thresholds method was suitable for the extraction of low,medium and high FVC,with an average E_F of 5.32%and RMSE of 3.20%.(3)Based on two periods of FVC data extracted from UAV images which were collected on June 30 and July 8,2020,The SL2P algorithm was directly verified and indirectly verified.Direct verification results showed that the accuracy of the SL2P algorithm increased with the increase of the number of UAV images in the pixel.When the number of UAV images in the pixel reached 8,the validation accuracy was the best,R~2 was 0.89 and 0.84,respectively,and RMSE was 0.02.But the FVCs were underestimated with Bias of-0.0164 and-0.0171,respectively.Secondly,NDVI,SAVI,RVI,DVI,GNDVI,and PSSRa derived from Sentinel-2 images were used to construct linear regression models to estimate FVC,and the SL2P algorithm was indirectly verified based on the estimated FVC.The results were consistent with the results of direct verification,but also were underestimated.(4)To solve the problem of underestimation of FVC for the SL2P algorithm,this study used the bands and vegetation indices derived from Sentinel-2 multispectral satellite data based on Random Forests(RF)and Gaussian Process Regression(GPR)to develop FVC estimation methods to improve the estimation accuracy.The estimation accuracy(R~2,RMSE)of the two methods was close to that of SL2P,but the Bias was reduced and the estimation Bias was between-0.0009 and 0.0007.