Deep Reinforcement Learning for Deficit Irrigation Scheduling in Olive Orchards: A field study based on the IoT in Semi-Arid

Abstract

The shortage of water in semi-arid regions of Iraq is one of the most serious restraints on agricultural production in the Arab World, while olive (Olea europaea L) is the plant that is best suited to the rapidly expanding dryland region of Iraq. Olive crops, which have been previously managed with fixed irrigation scheduling systems, not only waste water, but also do not match the expected yield. This research paper presents the first documented practice of an Internet of Things (IoT)-based irrigation control system using the Deep Q-Network (DQN) for Olea europaea in the extreme semi-arid climate of Baghdad (31.5N, 44.4E; July average Tmax 43.7C; average annual rainfall < 150 mm) by applying a phenomenological growth-stage weighted reward function for regulated water deficits during the synthesis of olive oil. As part of this study, each of the nine 80 m² research plot irrigation systems were equipped with a low-cost IOT platform (ESP32 + YL-69 + DHT22; USD 47.30 per zone) to function as the irrigation control system. In addition, the nine 80 m² research plot irrigation control systems were assigned in randomised complete block design to the following four separate irrigation control treatments. An olive-specific water balance model (using data from Baghdad Meteorological Station 2019–2024) provided a three-variable Markov Decision Process (soil moisture, air temperature, and relative humidity; 45 states; 3 actions) that was trained through 10,000 episodes (after which the sequence Kw showed that the algorithm converged at episode 1923). A systematic grid search of 81 configurations for the reward coefficient and a study of the ablation of 4 configurations to the state space yielded the first published evidence base for selecting design (DQN) choices for DQN developed to implement irrigation of perennial crops. In addition, three legitimate replicates at the plot level (June to August; Frantoio trees, 12 years old) were evaluated over 90 days of field evaluations. The data were analyses using a linear mixed-effects model, which corrected the pseudoreplication error of the past., and showed that DQN gave a 41.3% reduction in total seasonal water use as compared to control based on thresholds (LME analysis: F (3,267) = 61.84; p< 0.001 with all DQN contrasts, Tukey HSD; p<0.001), an oil yield of 97.1% of control (ANOVA: F (3,8) = 0.61; p=0.63; Non-significant), and a total Water Productivity Index of 0.803 kg/m³ (68.2% above sensor-based baseline).