Sahu et al. (2025) Optimizing Irrigation Scheduling with a Hybrid Transformer-GRU Model and Reinforcement Learning in Smart Agriculture

Identification

• Journal: Water Resources Management
• Year: 2025
• Date: 2025-12-29
• Authors: Shriya Sahu, Prerna Verma
• DOI: 10.1007/s11269-025-04461-7

Research Groups

• Department of Computer Science & Application, Atal Bihari Vajpayee Vishwavidyalaya, Bilaspur, Chhattisgarh, India

Short Summary

This research proposes an intelligent irrigation scheduling framework that integrates multi-source environmental data with a hybrid Transformer-GRU model and reinforcement learning to enhance water usage efficiency and crop yield while reducing drought risk in smart agriculture.

Objective

• To develop an intelligent irrigation scheduling framework that integrates multi-source environmental data with advanced deep learning (hybrid Transformer-GRU) and reinforcement learning techniques to optimize water application strategies, enhance water usage efficiency, and boost agricultural productivity.

Study Configuration

• Spatial Scale: Model-based study applicable to agricultural fields (e.g., maize farms). Future work aims for large-scale field deployment.
• Temporal Scale: The model captures long-range temporal dependencies and short-term sequential patterns. Evaluation metrics are presented over 5-day periods (soil moisture, NDVI, crop health) and monthly (drought risk frequency).

Methodology and Data

• Models used: Hybrid Transformer-GRU model, Reinforcement Learning (RL) agent, Autoencoder, Feedforward Neural Network (FNN).
• Data sources: IoT-based soil moisture sensor readings, weather station parameters (temperature, humidity, rainfall), remote-sensing vegetation indices (NDVI), static soil attributes (soil type), static crop attributes (crop characteristics).

Main Results

• Achieved a Mean Absolute Error (MAE) of 0.85 percentage points for soil moisture prediction.
• Improved water usage efficiency by 35% compared to traditional irrigation.
• Reduced irrigation volume to 0.65 mm (6,500 L/hectare) from a traditional 1.0 mm (10,000 L/hectare).
• Increased crop yield to 6,200 kg/hectare, outperforming traditional methods (4,500 kg/hectare) and drip irrigation (5,900 kg/hectare).
• Reduced drought risk by 60% (from 5 events per month to 2 events per month).
• Maintained stable average soil moisture levels around 27.5%.
• Enhanced Normalized Difference Vegetation Index (NDVI) by 0.10, indicating healthier vegetation.
• The reinforcement learning agent achieved optimal policy convergence with a reward value of 1.00.

Contributions

• Proposes a novel intelligent irrigation scheduling framework by uniquely integrating a hybrid Transformer-GRU deep learning model with a reinforcement learning agent.
• Demonstrates superior performance in prediction accuracy, water conservation, and crop productivity compared to traditional and existing machine learning models (Simple GRU, Simple Transformer, Random Forest, SVR).
• Introduces a sustainability-aware reward function for adaptive irrigation control, balancing water conservation with crop health.

Funding

• Not Applicable.

Citation

@article{Sahu2025Optimizing,
  author = {Sahu, Shriya and Verma, Prerna},
  title = {Optimizing Irrigation Scheduling with a Hybrid Transformer-GRU Model and Reinforcement Learning in Smart Agriculture},
  journal = {Water Resources Management},
  year = {2025},
  doi = {10.1007/s11269-025-04461-7},
  url = {https://doi.org/10.1007/s11269-025-04461-7}
}