Crop Yield Forecasting and Water Stress Assessment Using Sentinel-2 and Machine Learning

This paper analyzes the spatial distribution of vegetation status and moisture availability in agricultural crops based on multispectral satellite observations from the Sentinel-2 system across the Ashkadar agricultural production cooperative in the Sterlitamak District of the Republic of Bashkortostan, covering a total area of over 1000 hectares during the 2024–2025 growing season. The influence of spectral characteristics of vegetation cover on the actual yields of corn, wheat, and soybeans was examined, ranging from 42 to 578 centners per hectare with plant heights ranging from 79 to 272 centimeters. The presence of stable dependencies was revealed between the normalized difference vegetation index in the range from 0.70 to 0.92, the normalized difference moisture index in the range from 0.34 to 0.48 and the yield level in the range from 40 to 580 centners per hectare, while the Pearson correlation coefficient between these indices reaches 0.93. The effect of water deficit on intra-field yield heterogeneity was studied, showing that differences within a single field reach 100–150 centners per hectare when the normalized difference moisture index decreases below 0.36. A quantitative relationship was determined according to which, with a normalized difference moisture index of less than 0.36, crop yields decrease to 40–60 centners per hectare, whereas with index values above 0.44, crop yields exceed 400–500 centners per hectare, and the yield potential realization coefficient ranges from 0.62 to 0.85. An increase in crop yields in irrigated areas compared to non-irrigated ones was found by an average of 15–25%, and it was also discovered that up to 30–40% of the area of non-irrigated fields is subject to moderate and severe water stress. A set of input features was formed from six spectral channels and two integral indices used to model crop yields at the level of individual fields and calculate the yield potential realization coefficient. A regression solution based on an ensemble random forest algorithm was developed, yielding an average relative forecast error of less than 10% and a determination coefficient of 0.78, with a Pearson correlation coefficient between actual yield and the predicted yield potential of 0.58. The use of satellite remote sensing data and machine learning methods to generate yield forecast maps 2–3 months before harvesting and for the early identification of water-deficit zones is justified, enabling targeted irrigation management and the reduction of production risks.