Study On Rice Yield Prediction Model Based On Canopy Hyperspectral Remote Sensing

Timely,rapid,accurate,and non-destructive grain yield forecasting is of great significance for guiding agricultural production and formulating national grain policies.This study conducted a two factor split plot design experiment using different rice varieties and nitrogen application levels as experimental factors to determine the spectral reflectance of rice canopy at the jointing,booting,and heading stages,and collect agricultural parameter information.The hyperspectral characteristics of rice canopy at different nitrogen application levels and growth stages were analyzed,and the correlation between the original spectral reflectance,first-order derivative spectral reflectance,and yield was analyzed,as well as agricultural parameters The correlation analysis of vegetation index,characteristic parameters,yield and yield components was carried out.The regression method,BP neural network（BPNN）,support vector machine（SVM）and Random forest（RF）methods were used to construct the yield prediction models of one variable and multiple variables based on different independent variables,and the Coefficient of determination（R²）and Root-mean-square deviation（RMSE）were used to evaluate the prediction model.The main research findings are as follows:（1）Based on the hyperspectral characteristics of rice canopy at different nitrogen application levels,as the nitrogen application rate increases,the original hyperspectral reflectance of Qyou 6 and Yixiangyou 2115 shows a trend of first increasing and then decreasing,while Huanghuazhan shows an upward trend;For first-order derivative hyperspectral features,at 480-600 nm,Yixiangyou 2115 shows a trend of first increasing and then decreasing,while Qyou 6 and Huang Huazhan show opposite trends.At 680-780 nm,Qyou 6 and Yixiangyou 2115 showed a trend of first increasing and then decreasing,while Huang Huazhan showed an upward trend.Based on the hyperspectral characteristics of rice canopy at different growth stages,from the jointing stage to the heading stage,the original spectral reflectance of the three varieties showed an overall trend of first increase and then decrease,while the first derivative hyperspectral reflectance showed an overall trend of first decrease and then increase.（2）The correlation between the original hyperspectral reflectance and yield during three growth stages is negative between 401 and 720 nm,positive between 740and 1349 nm and 1520 to 1779 nm,and low between 1991 and 2399 nm;The correlation between first-order derivative hyperspectral reflectance and yield during the three growth stages shows a fluctuating trend overall.The correlation between vegetation index and yield and yield composition reached extremely significant levels in all three growth stages,and the correlation between most characteristic parameters and yield reached extremely significant levels in all three growth stages.The correlation between agricultural parameters and yield and yield composition reached extremely significant levels in both booting and heading stages,and the correlation was significantly higher in booting and heading stages than in jointing stage.（3）The rice yield prediction models constructed based on different independent variables and methods have differences.Overall,the accuracy of the models at booting and heading stages is better than that at jointing stage,as follows:The optimal univariate prediction models for booting and heading stages,with agronomic parameters as independent variables,are parabolic function models constructed based on aboveground nitrogen accumulation（APNA）,with the model expression y=-0.3×APNA²+90.21×APNA+3214.26 and y=-0.23×APNA²+81.04×APNA+3090.06,the R2 of the test set reached extremely significant levels,with values of 0.647 and 0.494,respectively,and RMSE of 884.36 kg/hm² and 1067.75kg/hm²,respectively;The optimal multivariate prediction models for both reproductive periods are SVM models,with R² of 0.564 and 0.549 for the test set,and RMSE of 967.25 kg/hm² and 1059.27 kg/hm²,respectively.The optimal univariate prediction models for booting and heading stages,with vegetation index as the independent variable,are logarithmic function models and linear function models constructed based on optimized soil regulated vegetation index（OSAVI）and normalized vegetation index（RDVI）,respectively.The model expressions are y=8094.57+2.25E+05×Ln(OSAVI_{（1342,1779）})and y=1.38E+05×RDVI_{（961,1746）}+1267.76,the R² of the test set were 0.742 and 0.733,respectively,reaching extremely significant levels,with RMSE of 797.93 kg/hm² and 806.16kg/hm²,respectively;The optimal multivariate prediction models for both reproductive periods are BPNN model,with R² of 0.695 and 0.761 for the test set,and RMSE of 1151.82 kg/hm² and 892.20 kg/hm²,respectively.The optimal univariate prediction models for booting and heading stages,with feature parameters as independent variables,are logarithmic function models constructed based on the normalized values of red edge area and blue edge area[（SDr-SDb）/（SDr+SDb）],with the expressions y=2.16E+04+1.23E+04,respectively×Ln[（SDr-SDb）/（SDr+SDb）]and y=2.05E+04+1.19E+04×Ln[（SDr SDb）/（SDr+SDb）],the R² of the test set reached extremely significant levels,0.618 and 0.622,respectively,with RMSE of 902.13 kg/hm² and 907.29 kg/hm²,respectively.The optimal multivariate prediction models for both reproductive periods are BPNN model,with R² of 0.683 and 0.504 for the test set,and RMSE of 900.45 kg/hm² and1083.30 kg/hm²,respectively.The multivariate linear prediction model has a good prediction effect using the path of"vegetation index-agricultural parameters-yield".The model expressions for booting and heading stages are y=3324.70+7E+05×NDVI_{（1672,1685）}+12.85×APNA and y=1061.74-6.91E+05×OSAVI_{（1631,1650）}+8.52×APNA.The R² of and test set were0.713 and 0.732,respectively,and the RMSE was 786.12 kg/hm² and 758.39 kg/hm²,respectively.In summary,the correlation between yield and its composition with vegetation index,characteristic parameters,and agronomic parameters is significantly higher at booting and heading stages than at jointing stages.In contrast,the univariate prediction model and multivariate prediction model constructed based on vegetation index have better performance,with better prediction results at booting and heading stages.