Neupane et al. (2025) “QuantumIrrigation” – a new quantum computing python package for irrigation demand assessment

Identification

• Journal: Smart Agricultural Technology
• Year: 2025
• Date: 2025-10-09
• Authors: Adarsha Neupane, Vidya Samadi
• DOI: 10.1016/j.atech.2025.101523

Research Groups

• Department of Agricultural Sciences, Clemson University, Clemson, SC, USA
• Artificial Intelligence Research Institute for Science and Engineering (AIRISE), School of Computing, Clemson University, Clemson, SC, USA
• Division of Civil Engineering, Department of Engineering, University of Cambridge, England, UK

Short Summary

This study introduces QuantumIrrigation, a Python package integrating quantum computing with deep learning for irrigation demand assessment by predicting crop reference evapotranspiration (ETo) and soil water tension. It demonstrates that while classical and quantum models perform similarly for ETo, hybrid classical-quantum models significantly enhance soil water tension prediction accuracy and computational efficiency.

Objective

• Evaluate the potential of standalone quantum models (Variational Quantum Circuit - VQC, Quantum Long Short-Term Memory - QLSTM) in capturing nonlinear, nonstationary relationships inherent in ETo and soil moisture dynamics (expressed as soil water tension).
• Develop classical-quantum hybrid architectures by integrating quantum models with deep learning approaches (hybrid LSTM-VQC, hybrid LSTM-QLSTM, hybrid PatchTST-VQC, and hybrid PatchTST-QLSTM).
• Benchmark the performance of these quantum and hybrid models against established time-series models, including classical Long Short-Term Memory (LSTM) networks and Patch Time-Series Transformer (PatchTST).
• Perform input sensitivity analyses to assess feature importance and elucidate the influence of key meteorological and soil parameters on classical versus quantum-driven models.

Study Configuration

• Spatial Scale: An experimental research farm (4 hectares) at the University of Georgia’s Stripling Irrigation Research Park (SIRP) in Camilla, Georgia, USA.
• Temporal Scale:
  • Daily meteorological data from January 1, 1998, to December 31, 2024, for ETo prediction.
  • 15-minute meteorological data and hourly soil water tension/temperature data from cotton growing seasons (May to October) between 2019 and 2023, for soil water tension prediction.
  • Look-back window: 7 days for ETo prediction, 24 hours for soil water tension prediction.
  • Prediction length: 1 (daily for ETo, hourly for soil water tension).

Methodology and Data

• Models used:
  • Classical Deep Learning: Long Short-Term Memory (LSTM), Patch Time-Series Transformer (PatchTST).
  • Standalone Quantum: Variational Quantum Circuit (VQC), Quantum Long Short-Term Memory (QLSTM).
  • Hybrid Classical-Quantum: hybrid LSTM-VQC, hybrid LSTM-QLSTM, hybrid PatchTST-VQC, hybrid PatchTST-QLSTM.
  • Reference Model: FAO-56 Penman-Monteith method (for ETo calculation and comparison).
• Data sources:
  • University of Georgia (UGA) Weather Network (daily and 15-minute meteorological data).
  • University of Georgia’s Stripling Irrigation Research Park (SIRP) (hourly soil water tension at 6, 12, and 18 inches depths, and hourly soil temperature).
  • Input variables for ETo: total solar radiation (Srad, MJ/m²/d), maximum/minimum air temperature (Tmax, Tmin, °C), average vapor pressure (VP, kPa), average dew point (Tdew, °C), maximum/minimum relative humidity (RHmax, RHmin, %), average wind speed (WS, m/s).
  • Input variables for soil water tension: hourly resampled meteorological variables (Srad, T, RH, Tdew, VP, WS, R), hourly soil temperature (ST), hourly resampled FAO-56 Penman Monteith estimated ETo, and 48 lagged hourly values of weighted-average soil water tension (ψ, kPa).
  • Data preprocessing: Forward linear interpolation for missing values, Min-Max Scaling for normalization.
  • Evaluation metrics: Mean Absolute Error (MAE), Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE). Mean-Squared Error (MSE) was used as the loss function.
  • Sensitivity analysis: Leave-One-Feature-Out (LOFO) method.
  • Software: QuantumIrrigation Python package, PyTorch, PennyLane, Optuna.

Main Results

• ETo Prediction:
  • All models, except VQC, achieved satisfactory performance (NSE ≥ 0.5). LSTM achieved the highest NSE (0.59), while PatchTST recorded the highest KGE (0.72). MAE values were comparable (0.72 to 0.73 mm/day).
  • VQC consistently underperformed (NSE 0.53, KGE 0.63, MAE 0.76 mm/day).
  • Hybrid models did not yield substantial improvements in predictive accuracy over standalone classical or QLSTM models for ETo.
  • Sensitivity analysis identified solar radiation as the most influential variable for ETo prediction, followed by wind speed.
• Soil Water Tension Prediction:
  • Hybrid LSTM-QLSTM achieved the best performance (NSE 0.97, KGE 0.93, MAE 1.30 kPa).
  • Hybrid LSTM-VQC closely followed (NSE 0.96, KGE 0.91, MAE 1.50 kPa). These two hybrid models significantly outperformed all other configurations.
  • Standalone PatchTST and VQC showed lower performance (NSE 0.88 for both, VQC with lowest KGE 0.84 and highest MAE 3.86 kPa).
  • Integrating quantum models into classical architectures significantly enhanced their ability to capture fine-scale temporal variability in soil moisture stress conditions.
  • Sensitivity analysis showed that lagged soil water tension values were the most influential predictors for all models except PatchTST, which was most affected by relative humidity.
• Computational Efficiency:
  • Classical models (LSTM, PatchTST) were the most computationally efficient (PatchTST: 5.41 s/epoch for ETo; LSTM: 10.76 s/epoch for soil water tension) due to GPU execution.
  • Quantum-based models (standalone and hybrid) incurred significantly higher computational costs due to CPU simulation and quantum overhead. QLSTM was the most expensive (1046.90 s/epoch for ETo, 6228.52 s/epoch for soil water tension).
  • Hybrid LSTM-VQC and hybrid LSTM-QLSTM demonstrated improved efficiency compared to standalone quantum models for soil water tension prediction, benefiting from processing distilled features from LSTM layers.

Contributions

• First application of Quantum Neural Networks (QNNs) for the simultaneous prediction of ETo and soil water tension.
• Development of "QuantumIrrigation," a novel Python package integrating classical deep learning, standalone quantum computing, and classical-quantum hybrid architectures for irrigation demand assessment.
• Comprehensive benchmarking of standalone quantum models (VQC, QLSTM) and four hybrid classical-quantum models against state-of-the-art deep learning models (LSTM, PatchTST).
• Demonstration of domain-specific advantages of quantum-classical hybrid modeling, particularly for soil water tension prediction, showing superior predictive accuracy and greater computational efficiency compared to standalone quantum models.
• Identification of key meteorological and soil parameters influencing ETo and soil water tension predictions through sensitivity analysis.
• Advancement of irrigation demand forecasting and support for more efficient, data-driven irrigation decision making.

Funding

• USDA-NIFA (grant # 2023000603)

Citation

@article{Neupane2025QuantumIrrigation,
  author = {Neupane, Adarsha and Samadi, Vidya},
  title = {“QuantumIrrigation” – a new quantum computing python package for irrigation demand assessment},
  journal = {Smart Agricultural Technology},
  year = {2025},
  doi = {10.1016/j.atech.2025.101523},
  url = {https://doi.org/10.1016/j.atech.2025.101523}
}