Liu et al. (2025) Deep Reinforcement Learning for irrigation optimization: Advantages, opportunities, and challenges

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-01
• Authors: Jiamei Liu, Fangle Chang, Jiahong Yang, Xinyi Jie, Caiyun Lu, Chao Wang, Lei Xie, Longhua Ma, Hongye Su
• DOI: 10.1016/j.agwat.2025.110030

Research Groups

• Ningbo Innovation Center, Zhejiang University, China
• NingboTech University, China
• College of Control Science and Engineering, Zhejiang University, China
• Polytechnic Institute, Zhejiang University, China
• College of Engineering, China Agricultural University, Beijing, China

Short Summary

This paper systematically reviews the applications of Deep Reinforcement Learning (DRL) in irrigation optimization, highlighting its strengths in handling dynamic, high-dimensional environmental data for adaptive and long-term strategies, while also identifying key challenges like data scarcity, model interpretability, and difficulties in field deployment.

Objective

• To systematically review the applications of Deep Reinforcement Learning (DRL) in irrigation optimization, covering both pre-trained environments based on crop growth simulators and dynamic environments driven by real-time sensors, and to discuss its advantages, opportunities, and challenges.

Study Configuration

• Spatial Scale: Applications range from specific crop fields (e.g., cotton, rice, almond, wheat, tomato, grape greenhouse) to large-scale distributed farms, addressing issues of spatial heterogeneity in water distribution.
• Temporal Scale: Irrigation decision-making is considered over various periods, including daily, throughout entire growing seasons, and for long-term water-saving goals, often leveraging real-time sensor data and weather forecasts.

Methodology and Data

• Models used: Deep Reinforcement Learning (DRL) algorithms (DQN, Double DQN, PPO, REINFORCE, TRPO, DDPG, SAC, A3C, A2C, SARSA(λ)), Deep Learning (DL), Machine Learning (ML), Support Vector Regression (SVR), k-means clustering, Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BLSTM), Artificial Neural Networks (ANN), Generative Adversarial Networks (GANs), Transformer models, Physics-Informed Neural Networks (PINNs). Crop growth simulators: CropGym, SIMPLE model, DSSAT, APSIM-Wheat, water balance model, AquaCrop model, Simu5G simulator.
• Data sources: Real-time sensor data (soil moisture, plant water potential, temperature, humidity, light), weather information (temperature, rainfall, humidity, solar radiation, weather forecasts), soil information (soil nutrient content, soil permeability), satellite imagery, historical data, simulated crop growth dynamics.

Main Results

• DRL demonstrates strong capabilities in learning adaptive and long-term irrigation strategies directly from high-dimensional environmental data, effectively managing dynamic and non-linear agricultural environments.
• Various DRL algorithms (e.g., DQN, PPO, DDPG, SAC, A3C) have been successfully applied to irrigation optimization, showing promise for multi-objective optimization, such as maximizing crop yield and water use efficiency.
• Case studies highlight DRL's application in both pre-trained environments utilizing crop growth simulators (e.g., DSSAT, APSIM) and dynamic environments driven by real-time sensor data.
• DRL-based approaches have shown significant improvements, including water savings ranging from 9.52% to 50% and yield increases from 3.5% to 31.4% compared to conventional irrigation methods.
• Key challenges identified include the high data dependency and scarcity of high-quality, high-frequency agricultural data, the "black box" nature of DRL models hindering interpretability, and difficulties in practical field deployment due to the "sim-to-real gap" (e.g., non-uniform water application, complex external factors).
• Future research directions propose developing digital twin platforms with generative AI, creating hybrid intelligent systems integrating DRL with Large Language Models (LLMs) and human expertise, optimizing multi-objective DRL models, and building safe and offline DRL frameworks to address current limitations.

Contributions

• Provides a systematic and comprehensive review of Deep Reinforcement Learning (DRL) applications in irrigation optimization, categorizing them by environment type (simulator-based versus sensor-based).
• Discusses the applicability, strengths, and weaknesses of various classic DRL algorithms within the context of smart irrigation management.
• Identifies and elaborates on critical constraints and challenges, including data dependency and scarcity, model complexity and interpretability, and difficulties in field deployment, which are crucial for the practical adoption of DRL in agriculture.
• Proposes concrete future research directions and potential solutions, such as developing digital twin platforms, hybrid intelligent systems, multi-objective DRL models, and safe/offline RL frameworks, to advance the field.
• Synthesizes the current state-of-the-art, offering a valuable resource for researchers and practitioners aiming to leverage DRL for sustainable agricultural development and efficient water resource management.

Funding

• National Natural Science Foundation of China (6247070970)
• Ningbo Science and Technology Innovation Yongjiang 2035 Key Technology Project (2024Z265)
• Ningbo Key research and development Plan

Citation

@article{Liu2025Deep,
  author = {Liu, Jiamei and Chang, Fangle and Yang, Jiahong and Jie, Xinyi and Lu, Caiyun and Wang, Chao and Xie, Lei and Ma, Longhua and Su, Hongye},
  title = {Deep Reinforcement Learning for irrigation optimization: Advantages, opportunities, and challenges},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110030},
  url = {https://doi.org/10.1016/j.agwat.2025.110030}
}