Multi-modal data fusion and deep ensemble learning for accurate crop yield prediction
Akshay Dagadu Yewle, Laman Mirzayeva, Oktay Karakuş
Abstract
This study introduces RicEns-Net , a novel deep ensemble model for rice yield prediction in the Mekong Delta region of Vietnam, integrating diverse data sources through multi-modal data fusion. The model leverages synthetic aperture radar (SAR), optical remote sensing (Sentinel-1/2/3) and meteorological measurements (surface temperature, rainfall) to improve prediction precision. A comprehensive feature selection reduced over 100 potential predictors to 15 key features across 5 data modalities, mitigating the “ curse of dimensionality ” where the initial field data were acquired through Ernst & Young’s (EY) Open Science Challenge 2023. RicEns-Net outperforms previous state-of-the-art models (including winners of the EY Open Science Challenge), achieving a mean absolute error (MAE) of 336 kg/Ha, roughly 5%–6% of the lowest regional yield, and a high R 2 , indicating robust predictive capability. These results underscore the benefit of deep ensembles in precision agriculture and demonstrate the potential of multi-modal data integration for more accurate crop yield forecasting.