Introduction: Bio-inspired optimization algorithms, including the Firefly Algorithm and Grey Wolf Optimizer, have attracted increasing interest for agricultural water management. However, applying stochastic optimization without first verifying the existence of a physically and operationally feasible improvement region may lead to apparently optimal but scientifically unsupported irrigation strategies. This study proposes a constraint-aware decision framework in which bio-inspired optimization is considered only after predictive modelling and deterministic feasibility screening.
Methods: The study used 2,072 daily NASA POWER environmental observations for the Doukkala–El Jadida region of Morocco, covering June 2020 to February 2026. The dataset includes climatic variables and model-derived root-zone soil wetness (GWETROOT), which is treated as a model-based environmental variable rather than an in situ sensor measurement. A Delta Temporal CNN was developed to predict next-day changes in GWETROOT. Temporal model selection was conducted using expanding-window validation, temporal-window sensitivity analysis, and multi-seed stability assessment. A 7-day input window was selected before evaluation on a post-development historical holdout. For irrigation management, physically constrained water-balance simulations were used to define a reference policy and determine whether a defensible improvement region existed before considering a hybrid Firefly–Grey Wolf Optimizer (FA-GWO). Deterministic, operational, and local policy grids were evaluated under water-balance, water-stress, terminal-storage, irrigation-volume, and event-count constraints.
Results: The selected 7-day temporal CNN achieved a development RMSE of 0.004245 ± 0.000027 and an R² of 0.988984 ± 0.000174 across the multi-seed evaluation. On the 312-sample post-development historical holdout, the model achieved an RMSE of 0.006169, compared with 0.006220 for linear regression. This marginal difference does not support a strong superiority claim. More importantly, an initial deterministic screening of 2,430 predefined irrigation-policy experiments identified no non-baseline feasible policy under the imposed constraints. Subsequent operational and fine local searches likewise failed to identify a defensible improvement over the reference policy. Consequently, stochastic FA-GWO optimization was deliberately not executed.
Conclusions: The results suggest that bio-inspired optimization should be preceded by explicit feasibility and constraint screening. In the investigated configuration, withholding FA-GWO was scientifically more defensible than forcing an apparent optimum. The proposed workflow provides a transparent and reproducible basis for future bio-inspired irrigation optimization as additional field measurements, operational variables, and forecast-aware decision parameters become available.