Integrating Multi-Source and Multi-Temporal Features for Winter Wheat Yield Estimation Using Vegetation Indices and Growth Indicators

Reliable estimation of winter wheat yield is critical to food system stability and farmland management. Integrating multi-spectral remote sensing data with agronomic parameters represents a primary strategy for improving yield estimation accuracy. However, existing research often overlooks parameters reflecting crop population structure and fails to account for dynamic shifts in the contributions of multidimensional agronomic variables across growth stages, thereby limiting prediction accuracy and model stability. To address these limitations, a winter wheat yield estimation model was developed. This model integrates multi-source and multi-temporal data, incorporates stem tiller density, a key population structure parameter, and accounts for dynamic variation across growth stages. Unmanned aerial vehicle multi-spectral images were collected at four key growth stages: jointing (stem elongation with detectable nodes), booting (flag leaf sheath swelling preceding heading), heading (spike emergence) and filling (grain filling with dry matter accumulation). Three growth indicators, stem tiller density, leaf area index and above-ground biomass, were measured. Two comprehensive growth indicators were derived using the coefficient of variation and the CRITIC weighting methods (CGICR). Correlation and feature importance analyses were used to identify sensitive vegetation indices (VIs), which were subsequently integrated with the comprehensive growth indicators. Single-stage, multi-source feature fusion and multi-temporal yield estimation models were established using the Kernel Extreme Learning Machine and its optimised algorithm using the Crested Porcupine Optimizer. The results showed the following: (1) among the individual growth stages, features from the filling stage achieved the highest prediction accuracy; (2) the fusion of multi-source features (VIs + CGICR) enhanced the prediction accuracy of the model, achieving a validation set R2 of 0.884 and a relative prediction deviation of 2.916 at the filling stage; and (3) the multi-temporal model further improved predictive performance, with the validation R2 reaching 0.920, indicating that information from different growth stages contributed complementarily to yield prediction and improved overall model performance. By contrast, the model exhibited relatively weak predictive capability at the early growth stages and was better-suited to early risk identification. Meanwhile, its generalisation ability under cross-regional and inter-annual conditions still requires further validation. Overall, integrating multi-source and multi-temporal data can enhance the precision and stability of predicting winter wheat yield, thereby facilitating precision agriculture management.