Timely crop yield forecasting remains a crucial component of global food security and agricultural resilience, particularly in the face of increasing climate variability. Here, we present a novel, high-resolution in-season yield prediction framework that integrates biweekly solar-induced chlorophyll fluorescence (SIF) time series from OCO-2 with neural network models to forecast county-level corn yields across the U.S. Corn Belt.
Using five growing seasons (2015–2020) and 210 corn-dominated counties, we applied sub-pixel SIF disaggregation to isolate crop-specific photosynthetic signals. A feed-forward neural network was trained using nested leave-one-year-out cross-validation, and input feature importance was ranked with the Boruta algorithm. Results show that mid-to-late season SIF observations (July–September), aligning with critical phenological stages, yield the highest predictive power. Models achieved R² values of up to 0.86 and RMSEs below 0.5 Ton/Ha, outperforming NDVI- and NIRv-based baselines.
This study highlights the physiological advantages of SIF in capturing real-time crop responses to climate and environmental stress. By using only remote sensing inputs, the framework provides a scalable, data-efficient alternative to weather- or model-based systems, with demonstrated potential for deployment in data-scarce regions.
