IrriTwin: Integrating reinforcement learning with soil water and crop models toward digital-twin-oriented irrigation scheduling
Abstract
Improving agricultural water-use efficiency under increasing water scarcity is a central challenge for sustainable agriculture. However, intelligent irrigation control is constrained by the high cost of trial-and-error interaction in real fields, while purely reinforcement-learning-based approaches have limited ability to explicitly represent soil water dynamics and crop physiological processes. To address these limitations, this study proposes IrriTwin, a digital-twin-oriented irrigation scheduling framework that couples the Soft Actor-Critic (SAC) algorithm with a soil water model and the GreenLab crop growth model, thereby establishing a model-in-the-loop virtual soil–crop–water environment for policy learning under water resource constraints. Experimental results show that, under the same total irrigation amount, the SAC-based strategy achieves higher fruit biomass than baseline methods. The learned policy exhibits dynamic adaptability to soil moisture conditions and crop phenological stages, effectively balancing vegetative growth and reproductive demands. Additional comparisons with reinforcement learning baselines further indicate that the proposed dual-action SAC achieves more stable convergence and higher final returns in the representative maize case. Stability analysis further shows that the strategy remains effective under simulated rainfall and irrigation-failure disturbances. Experiments on sunflower and maize demonstrate the effectiveness of IrriTwin. These results indicate that integrating reinforcement learning with mechanistic soil–crop–water models provides a promising foundation for digital-twin-enabled, precise, and water-efficient irrigation management.
Related
- Global meta-analysis and machine learning uncover region-specific salinity thresholds and risks for sustainable wheat production under saline irrigation
- DeepWaveRL: Self-Supervised Full Waveform Inversion via Reinforcement Learning
- DispSAC: Surface-wave dispersion inversion using soft actor-critic reinforcement learning
- AeroTrans: Hourly AOD retrieval over land from MSG-1/SEVIRI imagery integrating Transformer and transfer learning
- Meta reinforcement learning for robust quadruped locomotion under joint failures
BibTeX
@article{xu2026irritwin,
title = {{IrriTwin: Integrating reinforcement learning with soil water and crop models toward digital-twin-oriented irrigation scheduling}},
author = {Menghua Xu and Xiujuan Wang and Mengzhen Kang and Fangle Chang and Jing Hua and Haoyu Wang and Philippe de Reffye},
journal = {Journal of Hydrology},
year = {2026},
month = oct,
doi = {10.1016/j.jhydrol.2026.136034},
url = {https://hal.inrae.fr/hal-05743603v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - IrriTwin: Integrating reinforcement learning with soil water and crop models toward digital-twin-oriented irrigation scheduling
AU - Menghua Xu
AU - Xiujuan Wang
AU - Mengzhen Kang
AU - Fangle Chang
AU - Jing Hua
AU - Haoyu Wang
AU - Philippe de Reffye
PY - 2026
DA - 2026/10/01
JO - Journal of Hydrology
DO - 10.1016/j.jhydrol.2026.136034
UR - https://hal.inrae.fr/hal-05743603v1
AB - Improving agricultural water-use efficiency under increasing water scarcity is a central challenge for sustainable agriculture. However, intelligent irrigation control is constrained by the high cost of trial-and-error interaction in real fields, while purely reinforcement-learning-based approaches have limited ability to explicitly represent soil water dynamics and crop physiological processes. To address these limitations, this study proposes IrriTwin, a digital-twin-oriented irrigation scheduling framework that couples the Soft Actor-Critic (SAC) algorithm with a soil water model and the GreenLab crop growth model, thereby establishing a model-in-the-loop virtual soil–crop–water environment for policy learning under water resource constraints. Experimental results show that, under the same total irrigation amount, the SAC-based strategy achieves higher fruit biomass than baseline methods. The learned policy exhibits dynamic adaptability to soil moisture conditions and crop phenological stages, effectively balancing vegetative growth and reproductive demands. Additional comparisons with reinforcement learning baselines further indicate that the proposed dual-action SAC achieves more stable convergence and higher final returns in the representative maize case. Stability analysis further shows that the strategy remains effective under simulated rainfall and irrigation-failure disturbances. Experiments on sunflower and maize demonstrate the effectiveness of IrriTwin. These results indicate that integrating reinforcement learning with mechanistic soil–crop–water models provides a promising foundation for digital-twin-enabled, precise, and water-efficient irrigation management.
SN - 0022-1694
SN - 1879-2707
ER -