Prediction of field winter wheat yield using fewer parameters at middle growth stage by linear regression and the BP neural network method

Simple and reliable methods for wheat yield forecasting at the field scale are significant for farmers to formulate effective and timely field management and estimate economic revenue. In this study, the linear regression method (least absolute shrinkage and selection operator regression, LASSO) and BP neural network method (back propagation neural network, BPNN) were chosen for field yield prediction. The four variable types considered are climate, soil moisture, crop growth, and flag leaf photosynthesis. Crop-related data were collected in 12 winter wheat varieties that are widely grown in the North China Plain from 2018 to 2021. The same data set was used for both LASSO and BPNN models' development and validation. The results show that models based on LASSO method had the determination coefficient (R2) from 0.23 to 0.78, and the root mean square error (RMSE) from 548 to 1023 kg ha-1, while they were 0.66-0.94 and 302-684 kg ha-1 in models based on the BPNN method. This showed that the BPNN method performed better in field wheat yield prediction than the LASSO method. For BPNN method, using only a single variable at the early grain-filling stage had the similar prediction accuracy to using single or comprehensive variables in multiple growth periods, indicating that the field wheat yield can be predicted earlier by BPNN method with fewer indicators. Last, an optimal field wheat prediction model was developed by the BPNN method using crop growth variables (dry matter and leaf area index) at the early grain-filling stage, with the R2 and RMSE of 0.93 and 288 kg ha-1 in the model development stage, respectively, and correspondingly 0.86 and 481 kg ha-1 in the model validation stage. The significant point of this model is that the field wheat yield can be predicted with only two growth indicators approximately 35 days before harvest for 12 mostly cultivated wheat varieties.