Revealing nonlinear thresholds and interactions in maize yield and water productivity responses to straw incorporation via an integrated meta-analysis and explainable machine learning framework

Enhancing maize yield and crop water productivity (WPc) is a central goal of sustainable agricultural development in China, particularly under the combined pressures of climate change, water scarcity, and the ongoing transition toward greener farming systems. As a low-carbon practice that improves soil conditions and recycles organic resources, straw return exhibits highly variable effects due to heterogeneity in ecological zones, soil characteristics, and management regimes—yet the dominant drivers and nonlinear response patterns underlying these effects remain insufficiently understood. This study compiled 1094 field observations from major maize-growing regions in China (2000–2024) and developed an integrated framework that combines meta-analysis, Synthetic Minority Over-sampling Technique (SMOTE)-based data augmentation, machine learning prediction, and Shapley additive explanations (SHAP) analysis to quantify the overall effects of straw return and identify its dominant controlling factors and nonlinear interactions. Meta-analysis showed that straw return significantly increased maize yield by 7.15 % (95 % CI: 6.85–7.45 %) and WPc by 6.16 % (95 % CI: 5.93–6.39 %). The most pronounced benefits were observed under moderate rainfall (350–450 mm), low-fertility acidic soils (pH < 6.5; soil organic matter < 10 g·kg −1 ), and optimal planting density (5.25–6.75 ×10 4 plants·ha −1 ). The SMOTE algorithm, enhanced with a spatial perturbation strategy, improved the representativeness of high-response samples in underrepresented regions. Among the three machine learning models tested, the random forest (RF) model achieved the highest predictive accuracy for both yield (R 2 = 0.93) and WPc (R 2 = 0.91), outperforming support vector regression (SVR) and gradient boosting (GB). Further SHAP analysis identified soil pH, planting density, exchangeable potassium, and organic matter as dominant drivers of yield response, while WPc was more strongly influenced by annual sunshine duration, soil pH, planting density, and total nitrogen. These results reveal the complex, nonlinear, and context-dependent nature of straw return effects, emphasizing the coordinated roles of soil fertility improvement and radiation management in enhancing maize yield and water use efficiency. Overall, this study uncovers the dominant predictive drivers and nonlinear response relationships of straw return and establishes a robust data-driven framework for optimizing region-specific field management and improving agricultural water use efficiency. The findings provide actionable insights for evidence-based agricultural management and policy formulation, supporting sustainable intensification in arid and semi-arid agroecosystems globally. • Meta-machine learning revealed strong regional differences in straw return effects on maize yield and water productivity. • Oversampling with spatial perturbation improved prediction accuracy and robustness for yield and water productivity. • Interpretability analysis showed soil pH, planting density, and fertility as key drivers with nonlinear thresholds.