Abstract:Precise estimation of winter wheat yield at the field plot scale was considered to be of significant practical importance for agricultural yield management under large-scale farming conditions. Given that single vegetation indices were insufficient to comprehensively characterize crop growth status, the development of multi-parameter and multi-variable UAV-based remote sensing yield estimation methods was regarded as an emerging trend. An air-ground integrated decision-making framework was adopted, and winter wheat from the 2023—2024 growing season was selected as the research object. UAV multispectral imagery acquired during the heading and grain filling stages was utilized and combined with a limited amount of ground-measured leaf area index (LAI) as a crop morphological parameter to construct yield estimation models, including single-variable models ( vegetation index only), dual- variable models (vegetation index + texture features), and triple-variable models (vegetation index + texture features + LAI). Yield estimation modelling was conducted by using random forest ( RF), extreme gradient boosting ( XGBoost ), and support vector machine ( SVM ) algorithms, and the estimation performance of different models and feature combinations was comparatively analyzed. The results indicated that the RF algorithm based on the three-variable combination of vegetation index, texture features, and LAI achieved the highest modelling accuracy at both growth stages (heading stage: R2 = 0. 729, RMSE = 524. 475 kg / hm2, NRMSE = 11. 181% , RE = 3. 481% ; grain filling stage: R2 = 0. 779, RMSE = 479. 265 kg / hm2, NRMSE = 9. 736% , RE = 3. 205% ), and significantly outperformed the XGBoost and SVM algorithms. Based on the optimal model-feature combination, spatial distribution maps of winter wheat yield for the 2023—2024 and 2024—2025 seasons during the heading and grain filling stages were generated. The results revealed that yield estimation accuracy during the grain filling stage was generally higher than that during the heading stage. The average differences between estimated and measured yields during the grain filling stage for the two seasons were 334. 035 kg / hm2 and 284. 235 kg / hm2, respectively. SHAP analysis further indicated that LAI contributed substantially to the improvement of yield estimation accuracy across all growth stages. Overall, the findings demonstrated that the proposed air-ground integrated multivariate stepwise fusion decision-making approach enabled rapid and accurate estimation of winter wheat yield at the field plot scale, thereby providing effective technical support for smart agricultural management.
