Machine learning-based in-season crop yield forecasting using climate, geographic, and satellite data

Overview

This solution delivers in-season yield forecasts for winter wheat by quantifying the interaction of weather, geographic, and agronomic factors through modern machine learning methods. By capturing the sequence of crop development states and identifying extreme conditions early, the models provide actionable yield estimates before final harvest, enabling farmers, advisors, and agricultural stakeholders to make informed decisions on inputs, logistics, and risk management. The approach also offers a framework for assessing the long-term effects of climate change on crop productivity.

Technical specifications

• Data foundation: A large curated dataset of 65,000 samples from winter wheat variety and fertilization experiments across Germany (2001–2019), integrating agronomic, geographic, soil, and climatic parameters
• Modeling approach: Multiple modern machine learning methods trained to detect patterns in the interaction of geographic features and climatic influences on plant growth
• Forecast accuracy: Pre-harvest yield estimates with a mean error of 0.6–1.1 tons per hectare, with best-case scenarios reaching 0.6 tons
• Satellite enhancement: Integration of high-resolution hyperspectral satellite imagery to determine biomass growth, improving model accuracy to R² values of 0.67–0.75
• Application focus: Winter wheat yield forecasting, with potential extension to other crops and regions

Technology readiness level

The models have been developed and validated on extensive historical experimental data, demonstrating reliable yield prediction performance. The next phase involves field-testing the existing scientific models in real agricultural practice through partnerships with farmers, adapting the models to new data qualities, and incorporating high-resolution satellite imagery to further improve accuracy. A one-year validation period is planned, with a spin-off envisioned based on the outcomes.