Liu et al. (2025) Smart Irrigation Scheduling for Crop Production Using a Crop Model and Improved Deep Reinforcement Learning

Identification

• Journal: Agriculture
• Year: 2025
• Date: 2025-12-11
• Authors: Jiamei Liu, Fangle Chang, Xiujuan Wang, Mengzhen Kang, Caiyun Lu, Chao Wang, Shanshan Hu, Yanfeng Liu, Longhua Ma, Hongye Su
• DOI: 10.3390/agriculture15242569

Research Groups

• School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou, China
• School of Information Science and Engineering, NingboTech University, Ningbo, China
• Ningbo Global Innovation Center, Zhejiang University, Ningbo, China
• Institute of Automation, Chinese Academy of Sciences, Beijing, China
• College of Engineering, China Agricultural University, Beijing, China
• AeroDigital Research Institute, Hangzhou, China

Short Summary

This paper proposes an intelligent irrigation scheduling method, Temporal–Spatial Attention Soft Actor–Critic (TSA-SAC), which integrates a crop growth model (DSSAT) with an improved deep reinforcement learning agent to optimize cotton yield and water use efficiency. The method achieved a 39.0% improvement in water use efficiency and an 8.4% increase in yield compared to fixed-schedule irrigation strategies.

Objective

• To synergistically optimize crop yield and water use in arid regions through intelligent irrigation scheduling.
• To establish a high-fidelity virtual environment for deep reinforcement learning (DRL) training using a rigorously calibrated DSSAT model.
• To develop a novel temporal state representation module, incorporating a Bidirectional Long Short-Term Memory (BiLSTM) network and a dual-attention mechanism, to capture dynamic trends and identify salient information for agricultural decision-making.
• To introduce an innovative dynamic reward function based on the critical phenological stages of cotton, embedding agronomic prior knowledge into the optimization objective.

Study Configuration

• Spatial Scale: Shihezi City (44°18′ N, 85°59′ E), Xinjiang, China, a cotton cultivation area with a typical continental arid climate.
• Temporal Scale:
  • Meteorological data: 10-year period (2014–2024).
  • DRL training set: 7 years (2014–2020).
  • DRL validation set: 1 year (2021).
  • DRL test set: 3 years (2022–2024).
  • Daily time step for crop growth simulations and irrigation decisions.
  • Cotton growing season.

Methodology and Data

• Models used:
  • Crop growth model: Decision Support System for Agrotechnology Transfer (DSSAT v4.8) with the CSM-CROPGRO-Cotton module.
  • Deep Reinforcement Learning (DRL) agent: Temporal–Spatial Attention Soft Actor–Critic (TSA-SAC), an improved Soft Actor–Critic (SAC) algorithm.
  • Neural network components: Bidirectional Long Short-Term Memory (BiLSTM) network for temporal feature extraction, and dual-attention mechanisms (temporal and feature attention) for context aggregation and decision-making precision.
  • Reference evapotranspiration (ET0) calculation: Penman–Monteith method (FAO-56).
• Data sources:
  • Meteorological data: Daily records (maximum/minimum air temperature, precipitation, mean wind speed) from the China Meteorological Data Service Centre (2014–2024). Solar radiation estimated using the Angstrom–Prescott model.
  • Soil data: Stratified sampling and analysis from the agricultural experimental station of Shihezi University. Predominant soil type: Calcic Cambisol (WRB).
  • Crop management data: Simulated wide–narrow row configuration, in-row plant spacing, planting date (15 April), plant density (225,000 plants per hectare), and a split fertilization scheme (240 kg/ha nitrogen, 120 kg/ha phosphorus).
  • Field trial observation data: 2022–2024 for DSSAT model calibration and validation (aboveground biomass, seed cotton yield).

Main Results

• The proposed TSA-SAC method improved water use efficiency (WUE) by 39.0% compared to fixed-schedule irrigation strategies.
• TSA-SAC increased cotton yield by 8.4% (7421 kg/ha) and reduced irrigation water consumption by 22.1% (378 mm) compared to the Farmer Experience (FE) strategy.
• TSA-SAC achieved an irrigation water use efficiency (IWUE) of 19.6 kg/ha/mm, outperforming all benchmark DRL methods (DDPG, PPO, SAC, LSTM-SAC) and FE.
• The method significantly reduced cumulative water stress days (CWSD) from 25.3 days (FE) to 2.3 days, indicating effective prevention of detrimental water stress.
• An ablation study confirmed the significant contributions of the BiLSTM module (+9.3% IWUE), the dual-attention mechanism (+2.2% yield), and the dynamic reward function (+2.6% IWUE, +0.8% yield) to the overall performance.
• Model interpretability analysis showed dynamic attention shifts to critical features (e.g., root depth and soil water at seedling, leaf area index and biomass at squaring, water stress factor at flowering, reference evapotranspiration at boll-opening, and growth stage code at maturity) based on cotton's phenological stages.
• The agent demonstrated robustness, maintaining high performance (average yield decrease of only 1.4%) when 10% Gaussian noise was introduced to observation variables during inference.

Contributions

• Developed a robust, high-fidelity virtual environment for DRL agent training by encapsulating a rigorously calibrated DSSAT crop growth model with local data.
• Introduced a novel temporal state representation module combining a BiLSTM network and a dual-attention mechanism to effectively capture historical context and identify critical decision factors in agricultural environments.
• Designed an innovative dynamic reward function that integrates agronomic prior knowledge by adjusting weights based on cotton's critical phenological stages, enabling the DRL agent to balance yield maximization and water conservation.
• Demonstrated superior performance of the TSA-SAC framework over traditional and other DRL methods in optimizing cotton yield and water use efficiency, providing a significant advancement in smart irrigation scheduling.
• Enhanced the interpretability of DRL-based irrigation decisions through visual analysis of temporal and feature attention weights, increasing trustworthiness for practical application.

Funding

• National Natural Science Foundation of China, grant number 6247070970.
• Ningbo Science and Technology Innovation Yongjiang 2035 Key Technology Project, grant number 2024Z265.

Citation

@article{Liu2025Smart,
  author = {Liu, Jiamei and Chang, Fangle and Wang, Xiujuan and Kang, Mengzhen and Lu, Caiyun and Wang, Chao and Hu, Shanshan and Liu, Yanfeng and Ma, Longhua and Su, Hongye},
  title = {Smart Irrigation Scheduling for Crop Production Using a Crop Model and Improved Deep Reinforcement Learning},
  journal = {Agriculture},
  year = {2025},
  doi = {10.3390/agriculture15242569},
  url = {https://doi.org/10.3390/agriculture15242569}
}