Multi-modal representation for UAV-based rice yield prediction: Understanding temporal redundancy and feature sufficiency

In intensive heterogeneous smallholder farming systems such as the Vietnam Mekong Delta (VMD), high-precision and timely rice yield estimation is essential for effective agricultural decision-making. While UAV-based remote sensing provides valuable information, the role of temporal observations relative to feature representation remains unclear under heterogeneous field conditions. This study evaluates the ability of multi-modal feature representations to compensate for limited temporal observations. Spectral features, categorical agronomic variables, and deep spatial features were integrated and evaluated under both multi-temporal and single-temporal configurations. Feature contributions were analysed using SHAP values to characterise the temporal distribution and sufficiency of multi-modal representations. Results show that multi-modal feature integration consistently improves predictive performance across temporal settings. At the flowering (FL) stage, single-temporal models achieved high predictive accuracy ( R 2 > 0.80 ) approximately 30 days before harvest. While the highest predictive accuracy was observed at later growth stages, FL-stage models provided the best trade-off between prediction accuracy and operational lead time. SHAP analysis indicates that predictive information is predominantly captured by FL-stage features. Together with the comparable predictive performance observed between the single-temporal and multi-temporal configurations, these results suggest that additional temporal observations provide limited incremental predictive benefit within the evaluated observation window. These findings indicate that feature-rich representations at a critical phenological stage can compensate for reduced temporal aggregation, highlighting feature sufficiency as a crucial factor for data-efficient crop monitoring and enabling reliable early yield estimation in complex production environments.