Interpretable across-period ensemble learning with multi-stress dataset fusion to enhance early-stage yield prediction under combined water and nitrogen stress using hyperspectral sensing

Machine learning models fitted with remote sensing data facilitate accessible and timely crop yield prediction. Since crop canopy structure influences dry matter accumulation and signals physiological status, the predictive accuracy of dependent models improve with accumulating morphogenetic change, which limits their applicability in early-stage task. Moreover, interannual variations in combined stress (i.e., the simultaneous occurrence of multiple stressors) introduce generalization errors. which refer to the errors a model produces when applied to data that differs from its training set. Consequently, substantial research achieve limited practical adoption. This study takes winter wheat under combined water-nitrogen stress as the research object, conducted a two-year field trial both two experimental sites, utilizing raw canopy reflectance acquired through UAV-based hyperspectral sensing as modeling features. Quantification of model performance degradation under combined stress, revealed an average R² reduction of 13.6 % across Random Forest Regression (RFR), Light Gradient Boosting Machine, and Partial Least Squares Regression (PLSR) models at both sites. To this, models trained on the multi-stress dataset fusion strategy we proposed, achieved an average 8.51 % R² improvement compared to the best-performing single-stress dataset. Building upon the characteristic improvement of model accuracy with advancing growth periods, this study developed the Across-Period Ensemble Learning (APEL) framework. In heading-stage prediction across both regions, the APEL-PLSR models achieved R² of 0.72 and 0.647, outperforming its base learner PLSR models, while reducing RMSE by 78.2 % and 44.4 %. Compared to RFR, APEL-PLSR demonstrated an average R² increase of 0.389 and an RMSE reduction of 1.21 t/ha. We conducted Ablation Study on the APEL framework and integrated Variable Importance in the Projection and SHapley Additive exPlanations, further enhancing model interpretability. This study proposes a novel solution for early-stage yields prediction under complex stresses. • Quantified the detrimental effect of combined stress on yield prediction model. • Multi-stress dataset fusion enhances model robustness under combined stress. • Proposed a ensemble learning for accurate yield prediction before flowering. • Analyzed the impact of regional meteorological disparities on remote sensing. • Interpretability analysis of model structure and hyperspectral features.