A spatiotemporal convolutional neural network with explainable AI (XAI) unveils biofertilizer-driven ecophysiological mechanisms in maize

• An explainable Spatiotemporal CNN (ST-CNN) model was developed to predict maize yield, identifying key biofertilizer-driven features (AMF root colonization, leaf chlorophyll content, and plant N status) via SHAP and counterfactual analysis. • The ST-CNN framework decoded complex, nonlinear interactions between microbial inoculants and plant traits, providing a transparent AI tool for optimizing biofertilizer use in sustainable maize production systems. • Counterfactual simulations revealed actionable insights for precision agronomy, showing how incremental changes in root colonization and canopy temperature can maximize yield under climate variability. • The model outperformed traditional methods (Random Forest, Linear Regression), offering a robust, data-efficient pipeline for yield prediction and management recommendation even with limited datasets. • This study bridges the gap between microbial ecology and digital agriculture, delivering a smart, interpretable AI framework to enhance crop resilience and support decision-making for farmers and agronomists. The global demand for maize ( Zea mays L.) necessitates innovative approaches to enhance yield under increasingly adverse environmental conditions. This study quantifies the contributions of arbuscular mycorrhizal fungi (AMF) and plant growth-promoting rhizobacteria (PGPR) to maize grain yield using a multidisciplinary approach that integrates plant ecophysiology, soil ecology, and machine learning. Field experiments were conducted over two growing seasons under varying cropping systems and biofertilizer treatments. A spatiotemporal convolutional neural network (ST-CNN) was developed to predict maize yield from a comprehensive dataset encompassing 31 plant and soil features. Model interpretability was enhanced through SHapley Additive exPlanations (SHAP) analysis, which identified key traits such as leaf chlorophyll content (SPAD indices), plant nitrogen concentration, and root traits as primary determinants of yield. Counterfactual analysis further evaluated the effects of incremental changes in features, revealing nuanced nonlinear relationships, particularly highlighting the complex role of root colonization percentage by AMF. Although the ST-CNN model showed moderate predictive performance (R² = 0.459), it provided valuable insights into the ecophysiological mechanisms governing maize productivity and the synergistic effects of microbial inoculants. Comparative analyses with linear regression and Random Forest models underscored the challenges of modeling with limited datasets. The findings inform practical strategies for optimizing AMF and PGPR applications in maize production systems, advancing sustainable agriculture under climate change pressures. The integrated XAI framework unveils nonlinear and conditional mechanisms (e.g., threshold-dependent AMF effects) that complement traditional analyses, offering actionable insights for biofertilizer optimization.