Applying Machine Learning Models to Predict Population Dynamics of Harmful Non-Gregarious Locusts Across Kazakhstan Agroclimatic Zones

The paper reports the results of studies on the use of models based on machine learning algorithms and IT to predict the population dynamics of harmful, non-gregarious locusts in different agroclimatic zones of Kazakhstan. The research goal was to create a forecasting model for the population of harmful non-gregarious locusts in the agroclimatic zones of Kazakhstan using a machine learning algorithm and climatic predictors. The study covered the highly humid and slightly humid and moderately warm agroclimatic zones. Long-term data (2003-2023) on the number of pests and weather parameters (temperature, precipitation, soil water volume, etc.) were used to create multilinear regression, random forest, gradient boosting SVR, and SARIMA models. Model accuracy was assessed with MSE and R² metrics. As a result, Gradient Boosting SVR performed the best in terms of the accuracy and stability of forecasts for both zones. The key climatic parameters were determined: precipitation and soil water volume in July-August in the highly humid zone and spring soil water volume and early summer conditions in the slightly humid zone. Thus, together with GIS data, the highlighted model can make long-term phytosanitary forecasts and plan plant protection measures, considering regional agroclimatic features.