Use of a linear regression algorithm to predict the yield of “Oktava odeska” wheat in the steppe zone of mykolaiv region

Abstract

The article examines the application of machine learning models to forecast the yield of winter wheat variety “Octava Odeska” under the steppe conditions of the Mykolaiv region. It substantiates the need to replace traditional multifactor physical models with more flexible big data analysis algorithms. As input data, yield figures from 2020–2024, average winter and annual temperatures, and precipitation amounts were used, obtained from the open sources of the Ukrainian Hydrometeorological Center and Mykolaiv National Agrarian University. The linear regression and second-degree polynomial regression models were implemented in the Google Colaboratory Python environment, enabling rapid prototyping and validation of results using MAE and MSE metrics.Linear regression demonstrated stable additive rela- tionships: with each degree increase in average winter temperature, yield increases by approximately 0.08 t/ha, while each millimeter of winter precipitation reduces yield by about 0.06 t/ha (MAE = 0.47; MSE = 0.22). The fore- cast surface is flat, with a negative slope along the temperature axis and a positive slope along the precipitation axis, truncated by the genetic potential maximum of the variety (~ 5.7 t/ha).The second-degree polynomial model revealed pronounced nonlinear saturation and “plateau” effects: yield increases with moisture up to 50–60 mm and temperature up to + 2 … + 7 °C, followed by a slowdown in growth under extreme conditions (MAE = 0.0637; MSE = 0.0078). This approach better captures the interaction of climatic factors during critical phenophases.The derived relationships enable scenario-based yield forecasting for various climatic conditions, prompt planning of agronomic measures–irrigation, heat stress protection, adjustment of sowing dates–and bring actual results closer to the genetic potential of the “Octava Odeska” variety.Additionally, a sensitivity analysis of the model to input parameter deviations was conducted, confirming its robustness and practical applicability in variable climatic conditions.