Abstract
The object of the study is wheat yield forecasting based on the integration of climatic indicators, satellite vegetation indices, and machine learning algorithms. The problem to be solved is the limited accuracy of traditional crop yield forecasting methods, which fail to capture the complex nonlinear and multidimensional interactions among climatic, biophysical, and agronomic factors, thereby reducing their applicability for global food security tasks. The proposed approach is applied to a dataset comprising 345 observations from 2001–2023, combining vegetation indices (MODIS), climatic parameters (ERA5), and official statistics on yield and sown areas. The methodology included descriptive statistics, correlation analysis and forecasting models based on random forest, support vector machine and convolutional neural network. Model performance was assessed using coefficient of determination, root mean square error and mean absolute error. Random forest and support vector machine showed the highest accuracy (R2 = 0.85 with low errors), while convolutional neural network was less effective due to the limited dataset. The analysis confirmed the decisive role of vegetation indices, especially the normalized difference vegetation index, together with precipitation, temperature and sown area. The results address the identified research gap by demonstrating that the integration of climatic indicators and satellite vegetation indices significantly enhances the performance of machine learning models for wheat yield forecasting. In particular, the findings highlight the advantages of ensemble and support vector methods, which proved to be more robust and accurate under conditions of high climatic variability. The practical value lies in the potential use of these models in early warning and decision-support systems for farmers and state institutions, improving agrotechnical planning, resource allocation, and reducing food security risks, thereby contributing to global food security