Method for the prediction of wheat yield components

Accurate determination of yield components can assist in predicting the final crop yields, revealing the physiological significance of yield estimation. Research on the direct prediction of crop yield components is still lacking, because the feature data of yield components for long sequence are difficult to obtain, and some highly variable factors influence each other on the accuracy of the estimation. In this study, the spike quantity per plant (SQ), grain number per spike (GN), and 1000-grain weight (1 000 GW) of winter wheat were taken as prediction targets, to determine the optimal method for the prediction of winter wheat yield components. 126 independent factors were achieved using the puffing technology for meteorological factors after assessing the factors of planting species, density and region. A multivariable linear regression was used to analyze the crucial factors correlated to the concerned crop yield, and thereby to determine the quantitative relationship between the factors and yields. Three multiple regression models for the yield components of winter wheats were constructed after the 126 independent factors were regressed step by step. The determination coefficient R2 of the three multiple regression models were 0.515, 0.178 and 0.368, respectively, all at a low level than before. In collinearity diagnosis, if the characteristic values of multiple dimensions in 3 models were approaching to be zero, or the corresponding condition indexes were greater than 10, the time-delay prediction can occur due to the multicollinearity relation between factors. To solve this collinearity among factors and verify the data structure, a factor analysis was conducted to transform various observed variables into a few typical comprehensive factors. The optimized 126 independent variables made it possible to reduce the factor dimension. After factors optimization, 10 comprehensive factors were obtained to establish the three multiple regression predicting models of yield components, and the determination coefficient R2 were 0.376, 0.111 and 0.261, respectively, all less than 0.5. Based on neural network principle, a back-propagating neural network (BPNN) model was established between multiple independent factors and dependent variables, due to its ability for an approximate representation without restricting the input-output data. The determination coefficient R2 of the proposed model was 0.644 under the optimal model structure (126-16-3), indicating much better than that from the multiple linear regression and factor analysis. The overall prediction accuracy of BPNN model was 85.3%. The average prediction accuracies of grain number (GN) and 1000-grain weight (1 000 GW) were 88.1% and 89.5%, respectively, showing significantly higher than that of spike quantity per plant (SQ). In the prediction regions, the average prediction accuracies of the BPNN model were more than 80% in 6 regions, with the highest prediction accuracy of 89.6% in the east coast of Jiangsu. The results demonstrate that the nonlinear feature of BPNN model can be used to improve the approximation ability when dealing with multiple factors. The BPNN modeling is strongly recommended to predict yield components of winter wheat.