Machine learning for forecasting spring barley yield depending on agroclimatic conditions

Abstract

The research findings on forecasting spring barley yields depending on the agroclimatic conditions of the growing season using machine learning methods are discussed. To forecast barley yields, we used the data obtained from 2014 through 2024 in the field experiment of the Plant Gene Pool Laboratory of the Siberian Research Institute of Plant Production and Breeding (Branch of the Institute of Cytology and Genetics, Siberian Branch of the Russian Academy of Sciences) located in Krasnoobsk of the Novosibirsk Region. The agroclimatic resources during the years of research were calculated based on the data of the Ogurtsovo Agro-Meteorological Station. To train and test the machine learning models, we used three data sets containing various combinations of characters and the response variable - spring barley yield, g m². The following digital characters were used: average monthly and average ten-day air temperatures, °C; average monthly and average ten-day precipitation amounts, mm; accumulated air temperatures ≥ 10°C during the growing season, °C; total precipitation for the autumn-winter period (from September to March), mm; the duration of the growing season, days. The category characters were “variety” and “year of research”. Linear regression, linear regression with cross-validation, support vector machine for regression (SVR), random forest (RandomForest), and hyper-optimized gradient boosting (HGBoost) were used to construct models for predicting spring barley yields. The highest accuracy of crop yield forecasting was achieved when implementing the HGBoost model in all data sets. However, the highest accuracy (89.1%) was obtained in the data set containing average monthly air temperature and precipitation values, crop growing season duration, and the category characters “variety” and “year of research”. The other machine learning models were inferior to the HGBoost model regarding the accuracy of spring barley yield forecasts. The forecasting accuracy did not increase even when the data structure was changed. The modeling of spring barley yield forecasts showed that minimum amount of agroclimatic data was sufficient to achieve acceptable accuracy of the HGBoost model.