Research on County-Level Yield Simulation of Winter Wheat in Henan Province Based on Machine Learning Algorithms

LIU Xinglin; LIU Yuan; YANG Fan; LIU Buchun; HAN Rui

doi:10.15933/j.cnki.1004-3268.2025.08.017

Abstract：Henan is a major province for winter wheat cultivation,and simulating winter wheat yield is of great significance for ensuring national food security.This study analyzed the performance of machine learning models in winter wheat yield simulation using ten-day scale meteorological data and county-level winter wheat yield data from 16 counties(cities)in Henan Province from 2000 to 2019.The dataset was divided into a test set(2000-2015)and a training set(2016-2019).Based on multiple stepwise regression,random forest,and random forest OOB methods,county-level yield simulation models for winter wheat in Henan Province were constructed,and the simulation effects of different models were verified and compared.The results showed that,from 2000 to 2019,the winter wheat yield in Henan Province fluctuated between 2 001 and 7 980 kg/ha,with an average of 5 675 kg/ha and a coefficient of variation ranging from 3.75%to 26.58%.A multiple stepwise regression model was constructed based on 19 ten-day scale meteorological factors that passed the 95%significance test.The multiple stepwise regression model was validated with a determination coefficient(R2)of 0.620 9 and a root mean square error(RMSE)of 907.06 kg/ha;The random forest model constructed using all the characteristic factors was validated with the R2 of 0.772 5,and the RMSE of 664.36 kg/ha.A total of 68 key ten-day scale meteorological characteristic factors were screened based on random forest OOB importance analysis,among which,the ten-day scale meteorological factors in November last year,March,April and June had particularly significant impacts on winter wheat yield.The validation determination coefficient of the random forest OOB model for simulating county-level winter wheat yield was 0.860 5,and the RMSE was 636.58 kg/ha.The random forest OOB model performed better than the multiple stepwise regression model and the random forest model,with R2 increased by 38.59%and 11.39%,respectively,and RMSE decreased by 29.82%and 4.18%,respectively.This study utilized limited meteorological data and county-level yield data to achieve reliable and accurate winter wheat yield simulation,providing a methodological reference for regional winter wheat yield simulation.