The Role of AI Methods and Input Variable Selection in Time-Series Prediction for Smart Irrigation Systems

Abstract

Time series forecasting is very important for monitoring soil moisture and planning irrigation in agriculture. However, empirical studies on the impact of prediction algorithms, input parameters, and optimization strategies on predictive performance are still limited. The purpose of this study is to characterize the best prediction algorithms. The Systematic Literature Review (SLR) was PRISMA-based and consisted of 31 primary studies extracted from 99 publications from the Scopus database. The study evaluates the relationships between the prediction methods, input variables, and hybridization techniques used for model performance. The results indicate that deep learning (DL) and hybrid models were frequently associated with higher reported accuracy than classical machine learning (ML) within the reviewed studies; however, because performance figures were aggregated across heterogeneous metrics, datasets, and baselines, these tendencies are descriptive rather than pooled effect estimates. Inter-rater reliability for method categorization was almost perfect (Cohen’s kappa, κ = 0.895). The relevance of input variables, especially soil moisture, evapotranspiration, and remote sensing data, had a greater influence on accuracy than the number of features. This review also proposes a standardized benchmarking checklist to enable more comparable future evaluations.