Symmetric PSO Tuning of Deep Tabular Networks versus Grid-Search-Optimized Gradient Boosting for FAO-56 Surrogate Modeling of Reference Evapotranspiration under Progressive Climatic Input Scarcity: A Case Study in Bechar, Algeria

Abstract

Irrigated agriculture in arid regions have increased demand for accurate, low-cost estimation of reference evapotranspiration (ET0). Although the FAO-56 Penman–Monteith equation remains the reference method, its reliance on a complete meteorological record limits its applicability at automated stations with partial instrumentation, motivating machine-learning alternatives combined with metaheuristic tuning, such as Particle Swarm Optimization (PSO). However, comparative studies typically fine-tune only the proposed model, leaving baselines at their default settings. In order to close this gap, this study creates a transparent benchmarking protocol wherein Random Forest and XGBoost baselines are independently optimized via grid search, and the two deep tabular architectures, TabNet and Kolmogorov–Arnold Networks (KAN), receive an identical, symmetric Particle Swarm Optimization (PSO) budget. Daily meteorological data from an arid-zone station in Bechar, Algeria, are used to assess the models in three more enriched input scenarios: temperature alone, temperature and radiation, and the complete climatic set. Friedman-Nemenyi rank-based testing, ten random seeds, and a single chronological split were utilized in the primary comparison. Mean test R² rose from 0.843 to 0.894 to 0.993 across the three scenarios. Adding wind speed, wind direction, and relative humidity jointly reduced mean RMSE by 75%, compared with 18% for solar radiation alone; a feature ablation of three models attributed most of this gain to wind speed, with relative humidity contributing mainly in combination with it and wind direction giving no benefit. Friedman and Nemenyi tests ranked XGBoost first in S1 and S2. Under the full feature set, PSO-tuned TabNet matched XGBoost's average rank exactly (1.9 for both), and their pairwise comparison was non-significant (p = 1.000), indicating no detected difference in central tendency rather than demonstrated equivalence. These rank-based results are conditional on the seed set. An expanding-window diagnostic (one pass per fold) revealed that both top performing models gave strong performance with a mean R² value of 0.972 for XGBoost and 0.907 for the PSO tuned TabNet model. However, PSO tuning led to an obvious decrease in accuracy for the first fold compared to XGBoost (R² =0.663 vs. 0.952). These findings indicate that a grid-search-tuned gradient boosting model remains difficult to outperform under constrained inputs, and that metaheuristic tuning benefits deep tabular architectures conditionally rather than universally.