Enhanced wheat yield estimation with UAV multi-source remote sensing and optimized ensemble learning

• Multi-source data fusion enhances prediction accuracy:Combining UAV-derived RGB, multispectral data, and canopy height features achieves the highest yield prediction accuracy (SVR model with R²=0.673, RMSE=1106 kg/ha) during the winter wheat filling stage. • Ensemble learning strategy shows significant optimization:A selective ensemble learning approach using BPNN as the meta-model further improves accuracy to R²=0.711 and RMSE=1041 kg/ha, verifying the advantage of multi-model collaboration. • Key growth stage and feature screening:The filling stage is identified as the optimal prediction period. Red-edge and near-infrared vegetation indices (e.g., NNIR) exhibit the strongest correlation with yield (r=0.79), and canopy height features supplement 3D structural information. • Application value for precision agriculture:Provides a technical framework integrating multi-source data and ensemble learning for wheat yield estimation in semi-arid regions, supporting smart farm management decisions. Accurate prediction of crop yield is a critical issuefor smart field management. To address the limitations of single machine learning models (poor generalization) and single data sources (insufficient capacity to comprehensively characterize crop growth), this study acquired RGB and multispectral (MS) data of winter wheat across three key growth stages via a low-altitude UAV. After extracting vegetation indices (VIs) and crop height (CH) features, and combining them with measured yield data from 123 samples, it systematically evaluated the performance of six machine learning models and explored different combinations of ensemble learning strategies.The results showed that the ​filling stage​ was the optimal period for yield prediction, with ​Support Vector Regression (SVR)​​ achieving the highest accuracy (R²=0.673), outperforming the ​heading stage​ (R²=0.618) and ​milking stage​ (R²=0.534). ​Multi-source data fusion (MS+RGB+CH)​​ significantly improved prediction accuracy (p<0.05), with SVR performing best (R²=0.673, RMSE=1106 kg/ha, RRMSE=11.1%). Ensemble learning​ further enhanced performance when using a ​Backpropagation Neural Network (BPNN)​​ as the secondary model, increasing accuracy to ​R²=0.711, RMSE=1041 kg/ha, RRMSE=10.4%​. ​Selective removal of underperforming base models​ improved ensemble effectiveness. While BPNN performed best as a secondary model, Random Forest Regression (RFR) exhibited higher stability in this role. This study demonstrates that ​high-precision wheat yield prediction​ can be achieved through ​multi-sensor data fusion and selective ensemble learning, providing essential data support for smart agricultural management