European Journal of Sustainable Agroecosystems

Background: Background: Rainfed agricultural systems support nearly 80% of global cultivated land and provide the primary livelihood base for hundreds of millions of smallholder farmers, yet they remain highly exposed to interannual climate variability, making yield outcomes difficult to anticipate with conventional statistical and crop-simulation approaches. Over the past decade, machine learning (ML) and deep learning (DL) techniques have been increasingly applied to model the complex, nonlinear interactions among climatic, edaphic, and management variables that drive yield variability in these systems.
Objective: This review synthesizes peer-reviewed literature published between 2015 and 2026 on the application of ML and DL methods for predicting yield variability in rainfed cropping systems, with the aim of identifying dominant methodological trends, comparing model performance across contexts, and articulating unresolved research gaps.
Sources of literature: A structured search of Scopus, Web of Science, PubMed/PMC, IEEE Xplore, ScienceDirect, and Google Scholar, supplemented by reports from the Food and Agriculture Organization (FAO) and related international bodies, yielded 612 initial records; after de-duplication, title/abstract screening, and full-text eligibility assessment following a PRISMA-style process, 32 studies were retained for thematic synthesis.
Major findings: Ensemble tree-based algorithms, particularly Random Forest and Gradient Boosting/XGBoost, remain the most frequently reported and most robust models under data-scarce, high-variability rainfed conditions, while deep learning architectures (CNN, LSTM, and CNN-LSTM hybrids) achieve superior accuracy when large, temporally rich remote-sensing datasets are available. Rainfall, temperature, soil moisture, and vegetation indices (NDVI/EVI) consistently emerge as the dominant predictors of yield variability. Explainable AI (XAI) techniques such as SHAP and LIME are increasingly integrated to address the interpretability limitations of black-box models, and hybrid physics-informed or knowledge-guided ML frameworks are emerging as a bridge between process-based crop models and purely data-driven approaches.
Research gaps: Persistent gaps include scarcity of ground-truth yield data in smallholder and low-income regions, limited cross-regional transferability of trained models, weak capacity to predict extreme yield losses associated with compound climatic hazards, insufficient uncertainty quantification, and underrepresentation of Sub-Saharan Africa and South Asia relative to North America in the published evidence base.
Conclusion: ML-based approaches offer measurable improvements over conventional yield-forecasting methods in rainfed agriculture, but their operational value for smallholder decision-making and policy planning will depend on closing data, interpretability, and generalizability gaps through federated data-sharing, hybrid modelling, and participatory validation with end users.