Hybrid Explainable Artificial Intelligence and Evolutionary Optimization for Climate-Resilient Wheat and Maize Yield Prediction: A Systematic Review

Background: The prediction of crop yield, accurate and interpretable, is becoming more and more relevant to the quest of making agriculture climate resilient. The yield of wheat (Triticum aestivum L.) and maize (Zea mays L.) is influenced by the interaction of nonlinearity between climate change, soil moisture, crop growth, remote sensing signals and management. Machine-learning and deep-learning models can capture such complexity, but they are not very interpretable, and their extrapolation is unstable, and have been plagued by model-selection problems, limiting their use under changing climates.

Objective: To assess, in a systematic way, the wheat and maize yield prediction models that include explainable artificial intelligence (XAI) and/or evolutionary optimization, focusing on their predictive accuracy and transparency, optimization performance, robustness and generalizability under climatic and environmental variability.

Methods: The review was conducted according to the guidelines of the PRISMA 2020. A literature search was conducted in PubMed/MEDLINE, Scopus, Web of Science Core Collection, ScienceDirect, Emerald Insight and Google Scholar (including additional searching), up to 18 August 2026. A total of 155 empirical English-language studies were included, published between January 2016 and August 2026, that assessed wheat and/or maize yield as a result of applying machine learning/deep learning, with XAI, interpretable modelling, evolutionary optimization or a directly relevant hybrid approach. A narrative synthesis was conducted since study designs and metrics were diverse. Domains adopted from PROBAST+AI were used to assess the quality and applicability.

Results. A total of 2338 records were found. Following removal of 798 duplicates, 1540 titles/abstracts were screened with 189 full text reports assessed and 14 studies were included in the qualitative synthesis. The most popular XAI technique was SHapley Additive exPlanations (SHAP), and the most popular evolutionary optimizers were those based on genetic algorithms and NSGA-III. While XAI was consistently improving transparency, evolutionary optimization was also sometimes better than a good conventional tuning in feature selection or model configuration, but not always better than the conventional tuning. More evidence existed for external climate generalizability than there was for predictive accuracy.

Conclusions. The use of hybrid XAI and evolutionary optimization is a promising approach for wheat and maize yield prediction, but an R² or a low prediction error level is not an indicator of climate resilience. Stronger evidence necessitates temporally and geographically independent validation and evaluation under climatic extremes, uncertainty reporting, comparison with well-tuned non-evolutionary baselines, and assessment of the stability of XAI explanations in the face of environmental change, in an agronomically plausible sense.