Using remotely sensed vegetation indices and multi-stream deep learning improves county-level corn yield predictions

Abstract

The increasing global population and climate change pose significant threats to global food security. Machine learning models that use crop growth variables retrieved from remote sensing data offer the potential for early estimates of county-level crop yields, aiding informed decision-making for grain policy, insurance, and food security planning. While vegetation indices and other metrics derived from remote sensing data have been utilized in machine learning models for county-level crop yield estimation, multi-stream data fusion through deep learning remains underexplored. In this study, we trained random forest regression (RFR), support vector regression (SVR), deep neural network (DNN) and a multi-stream deep neural network (MSDNN) using vegetation indices (Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)) derived from Moderate Resolution Imaging Spectroradiometer (MODIS) to estimate county-level corn yield in Iowa state, within U.S. corn belt. For MSDNN, NDVI and EVI were used as two separate data streams initially to train the model. Additionally, we evaluated the effectiveness of early-season remote sensing data by dividing the growing season into four intervals: end of May (EoM), end of July (EoJ), and end of September (EoS). As compared to SVR, RFR and DNN, MSDNN model performed better with coefficient of determination (R²) = 0.86. For early season yield estimation, the EoS model yielded the highest accuracy (R² = 0.86), followed by EoJ (R² = 0.66) and EoM (R² = 0.35). These findings indicate that county-level corn yield can be reasonably estimated before harvest, offering valuable insights for informed grain policy decisions.