Research On Inversion Method Of Cotton Yield Based On UAV Visible Light Images

Due to the tedious planting process,long growth cycle and low degree mechanization of cotton,field management of cotton mostly relies on manual labor.The cotton planting area in some major cotton-growing provinces of Yellow River Basin has experienced a significant decline.With the rapid development of remote sensing,satellite positioning systems and geographic information systems,UAV remote sensing technology has been widely used.The low-altitude UAV remote sensing system provides technical support for small and medium-scale agricultural remote sensing monitoring due to its advantages of light weight,fast speed,high image resolution and short cycle.The Yu 2pro UAV was used to obtain the visible light images of the first bud stage,the full bud stage,the flowering stage and the boll stage of the cotton,and inverse prediction model of cotton yield was established by using above parameters.The research results provided a reference for the daily management of field cotton.(1)The 11 cotton varieties independently cultivated by the Shandong Cotton Research Center were used as the research object.The visible light images of the bare land,the first bud stage,the full bud stage,the flowering stage and the boll stage of the cotton were acquired by the Yu 2pro drone.Then,the Agisoft Photo Scan Professiona software was used to stitch the visible light images of the research field acquired by the UAV to generate DOM and DSM.(2)The blue-green characteristic vegetation index(TBVI)and red-green characteristic vegetation index(TRVI)were constructed by analyzing the spectral differences of the visible light bands of vegetation,soil,and shadows in the bud stage of cotton.Using TBVI,TRVI,Excessive Green Index(EXG),Selective Difference Vegetation Index(VDVI),Normalized Green-Blue Difference Index(NGBDI)and Red-Green Index(GRVI),combined with vegetation index threshold method,empirical model method and the maximum entropy threshold method,the vegetation coverage information of cotton in the full bud,florescence and boll stages were extracted.The support vector machine classification result was used as the true value.The cotton vegetation coverage extracted by the above three methods was verified by R~2,RMSE and n RMSE.The results showed that,compared with the maximum entropy threshold method and the empirical model method,the vegetation index threshold method had the best extraction effect.In the comparison of the vegetation index,vegetation coverage overall extraction effect of TRVI was the best(first buding period:R~2=0.8484,RMSE=2.0716,n RMSE=3.9648%;full bud period:R~2=0.8088,RMSE=2.5966,n RMSE=3.6329%,florescence period:R~2=0.9942,RMSE=0.5113,n RMSE=0.5903%,bell period:R~2=0.9775,RMSE=0.7792,n RMSE=0.8504%).According to the above evaluation results of vegetation coverage accuracy,the TRVI combined with vegetation index threshold method was used to extract the vegetation coverage in the four periods of cotton.The vegetation coverage in the four periods of cotton was 29.05%,55.96%,78.70%,and 91.74%,respectively.(3)In this paper,the DSM of cotton bare land was used as a benchmark,and the plant height parameters of cotton at the early bud,full bud and flowering stages were obtained by DSM difference method.The cotton plant height extraction effect was determined by observing the plant height gray map and plant height statistical histogram,and the sample area was selected in the cotton DSM image.The verification results showed that the DSM difference method had a better effect on vegetation coverage during the first bud,full bud and flowering stages(first bud period:R~2=0.781,RMSE=1.770,n RMSE=4.722%,full bud period:R~2=0.79,RMSE=2.994,n RMSE=4.449%,florescence period:R~2=0.8373,RMSE=4.0179,n RMSE=4.7101%).A vector file of cotton vegetation at the first bud,full bud and florescence stages extracted by TRVI combined with the vegetation index threshold method was generated,and it was applied to the parameters of cotton plant height,full bloom stage and flowering stage obtained by DSM difference method.Finally,the plant height distribution maps of cotton at the first bud and full bud stages were generated.Cotton plant height distribution map can facilitate field managers to understand the growth status of cotton and provide guidance for subsequent management.(4)This paper uses R~2 and RMSE as selection criteria for cotton yield inversion factors,and screens the visible light band vegetation index,plant height,texture characteristics,and vegetation coverage.The disparity of the red band in the 7×7 window and the information entropy of the red band in the 13×13 window in the visible light image at the flowering stage of cotton were selected as the inversion factors of cotton yield model.Combined with stepwise regression,partial least squares method and BP neural network method,the inversion prediction model of cotton yield was constructed,and the cotton yield prediction model was used to verify the cotton yield prediction model established previously.The results showed that the cotton yield inversion model based on the BP neural network model had higher accuracy(R~2=0.8533,RMSE=6.8826,n RMSE=8.4906%).