Spring Wheat Yield Forecast Using Machine Learning Methods

Abstract

AbstractThis paper presents the possibility and feasibility of using the Bayesian network method and multinomial logistic regression to predict the yield of spring wheat. To build and train the model the data of a long-term multifactorial stationary field experiment of the Siberian Research Institute of Husbandry and Chemicalization of Agriculture (SRIHCA) of the Siberian Federal Scientific Centre of Agro-BioTechnologies (SFSCA) of the Russian Academy of Sciences (RAS) for the time period of 2004–2018 were used. During the analysis of the data sample, the main predictors of the model affecting the spring wheat yield were identified. The predictors are represented by qualitative and quantitative parameters of the working area: predecessor, tillage, HTC (weather conditions), pesticides, and yield by appropriate gradations (events). As a result, models were built and tested that were able to predict the yield of spring wheat, depending on the prevailing conditions. To assess the predictive ability of the models were tested on the original sample. In Bayesian Networks, the total share of correct forecasts for all categories of spring wheat yields is 81%. The total share of correct forecasts obtained by implementing the multinomial regression model is 83%. The constructed models make it possible to predict with acceptable reliability.KeywordsGraphical modelRegressionPredictorsSpring wheatYieldForecast