Comparative analysis of reinforcement learning and rule-based system approaches for irrigation in vegetable production

Abstract

Efficient irrigation management is essential for horticulture in the face of climate variability and the need to optimize water use. This study compares irrigation control strategies based on Reinforcement Learning (RL), Rule-Based Systems (RBS), and a Hybrid model applied to lettuce cultivation. The approaches were evaluated in terms of crop productivity, wateruse efficiency, and adaptation to different climatic conditions. The proposed methodology integrated the AquaCrop-OSPy model as a simulation environment for training and evaluating the irrigation strategies, combined with artificial intelligence techniques and cloud computing. An RL model based on the Proximal Policy Optimization (PPO) algorithm was implemented, together with an RBS founded on agronomic variables such as soil moisture, evapotranspiration, phenological stage, and precipitation. The Hybrid model combined the adaptive capability of RL with the operational stability of the RBS. The experiments were conducted over four independent cultivation cycles using real meteorological data under different climatic conditions. The results revealed differences among the irrigation strategies regarding irrigation scheduling, response to rainfall events, and sensitivity to weather forecast uncertainty. In the first cycle, the irrigation depths were similar among the methods, reaching 75.66 mm for RBS, 75.26 mm for RL, and 73.80 mm for the Hybrid model, resulting in dry yields of 6.09 t/ha, 6.48 t/ha, and 6.59 t/ha, respectively. In the second cycle, irrigation depths increased to 124.40 mm (RBS), 149.14 mm (RL), and 131.60 mm (Hybrid), producing dry yields of 6.55 t/ha, 7.00 t/ha, and 7.52 t/ha, respectively. The greatest contrast occurred during the third cultivation cycle, characterized by frequent rainfall forecast errors. The RL strategy applied only 39.66 mm of irrigation, resulting in a dry yield of 4.52 t/ha. In comparison, the RBS applied 89.75 mm and achieved 7.48 t/ha, while the Hybrid model applied 112.75 mm and obtained the highest dry yield of the cycle, reaching 8.19 t/ha. These results indicate that excessive reductions in irrigation based on rainfall forecasts that do not materialize may negatively affect crop productivity. In the fourth cycle, irrigation depths reached 55.34 mm for the RBS, 84.24 mm for the RL, and 62.47 mm for the Hybrid model. Despite these differences, dry yields were nearly identical among the three strategies, reaching approximately 10.30 t/ha. This result indicates that higher irrigation depths do not necessarily lead to greater productivity, highlighting the importance of applying water at the appropriate time according to environmental conditions. Considering all four cultivation cycles, the Hybrid model achieved the highest overall performance, with an average dry yield of 8.15 t/ha, compared with 7.61 t/ha for the RBS and 7.32 t/ha for the RL. In addition, the Hybrid model demonstrated greater resilience to climatic uncertainty and weather forecast errors by combining the adaptive capability of Reinforcement Learning with the operational reliability provided by agronomic rules. The computational analysis also showed that the RBS has a low computational cost, whereas the RL approach requires greater processing capacity due to the learning process and continuous policy evaluation. The results indicate that integrating Reinforcement Learning with Rule-Based Systems is a promising approach for intelligent irrigation management, providing improved operational stability, adaptation to climatic conditions, and high agricultural productivity.