Feature sensitivity and interpretability in AI-driven rice yield prediction: A systematic review

Abstract

Paddy yield prediction has become increasingly critical in the context of global food security and climate variability. Artificial-Intelligence (AI) and machine learning (ML) provide powerful tools for improving prediction accuracy, but selecting and interpreting input features remain challenging due to complex agro-environmental interactions. This review aimed to (1) categorize critical features used in AI-based models, (2) assess the influence of various parameters on paddy yield predictions and (3) evaluate methodologies applied in feature selection and sensitivity analysis. A structured review of 89 peer-reviewed studies from Google Scholar was conducted using PRISMA guidelines, with data extracted to address four research questions. The review found that the most commonly used features fell into eight categories: climatic, soil, crop phenotypic, remote sensing, management, geospatial, temporal and stress/environmental factors. Rainfall and temperature were the most frequently used meteorological inputs, appearing in over 90% of studies. Evapotranspiration and cumulative rainfall were especially impactful in water-stress contexts. The importance of meteorological features varied by crop season, region, and irrigation practices, rainfall dominated in rain-fed systems, while temperature or vegetation indices were more influential in irrigated settings or during later crop stages. A wide range of feature selection and sensitivity analysis techniques was applied, including correlation-based methods (Pearson, Spearman), statistical techniques (stepwise regression, t tests, RFE, GAM), and model-intrinsic scoring tools (Gini index, SHAP, attention weights). Performance impact methods such as LOOCV, feature shuffling and ablation studies were also used, along with dimensionality reduction (PCA) and optimization algorithms (GA, PSO, SCA). Key challenges included high dimensionality, feature redundancy, temporal-spatial variability, poor data quality, lack of model interpretability and inconsistent reporting. Best practices identified included feature pre-screening (PCA), adopting temporally aware models like LSTM, applying explainable AI tools (SHAP, LIME, PDP), combining expert judgment with algorithmic selection, and ensuring standardized, interpretable reporting for better reproducibility and decision making in sustainable agriculture. These insights underscore the need for integrative, explainable and context-specific AI approaches to enhance the reliability of rice yield forecasting and support evidence-based decision for climate-resilient, sustainable agriculture.