Li et al. (2026) Research on inversion and prediction of root region soil water content in kiwifruit based on hyperparameter tuning by transformer-DsaGRU

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-01-06
• Authors: X M Li, Zili Chen, Jingyuan He, Qingyuan Liu, Zhen Niu, Zhilong Gao, Zefeng Jia, zijie niu, Dongyan Zhang, Mingu Zhou
• DOI: 10.1016/j.agwat.2025.110080

Research Groups

College of Mechanical and Electronic Engineering, Northwest A&F University, Yangling, China

Short Summary

This study developed a Transformer–DsaGRU model, integrating multi-source data and a rainfall-threshold-based dynamic step-size adjustment mechanism, to accurately forecast kiwifruit root zone soil water content (RSWC) 1-2 days ahead, demonstrating superior performance over traditional deep learning models.

Objective

• To develop and validate a Transformer-DsaGRU model for accurate 1-2 day ahead prediction of kiwifruit root region soil water content (RSWC) by fusing unmanned aerial vehicle (UAV) multispectral indices, meteorological variables, and a rainfall-threshold-based dynamic step-size adjustment mechanism.

Study Configuration

• Spatial Scale: A kiwifruit orchard covering 800 m² in Meixian County, Shaanxi Province, China (34°07′N, 107°59′E). Measurements were taken from 54 kiwifruit vines, with soil moisture sensors installed at a depth of 0.4 m.
• Temporal Scale: Field campaigns were conducted over three growing seasons: 7 June to 5 August 2023, 12 July to 22 August 2024, and 8 May to 20 August 2025. Data were collected daily in 2023 and 2024, and once every four days in 2025, focusing on the kiwifruit fruit expansion period. The model uses a base input window of 10 days, dynamically extended to 12 days under specific rainfall conditions, for 1- and 2-day ahead predictions.

Methodology and Data

• Models used:
  • Proposed: Transformer-DsaGRU (Transformer encoder + Dynamic step-size adjustment (Dsa) + Gated Recurrent Unit (GRU))
  • Baselines: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), GRU, Temporal Convolutional Network (TCN), Transformer-LSTM, Transformer-GRU.
  • Optimization: Bayesian optimization, random search, and grid search for hyperparameter tuning.
  • Feature selection: Pearson correlation (p < 0.05) and mutual information.
  • Feature contribution analysis: Random Forest–SHAP framework.
• Data sources:
  • Root Region Soil Water Content (RSWC): Measured using SN-3000-TR-N01 soil temperature–moisture sensors (accuracy ±2 % v/v, range 0–100 % v/v) at 0.4 m depth, calibrated with the gravimetric oven-drying method.
  • UAV Multispectral Imagery: Acquired using a DJI Mavic 3 Multispectral drone equipped with an RTK module. Cameras captured green (560 ± 16 nm), red (650 ± 16 nm), red-edge (730 ± 16 nm), and near-infrared (860 ± 26 nm) bands. Flights were conducted at 15 m altitude and 5 m/s speed. 15 vegetation indices (e.g., NDVI, SAVI, RTVICore, GCVI, OSAVI, gNDVI) were extracted.
  • Meteorological Data: Collected from a miniature automatic weather station at 15-minute intervals and aggregated to daily averages. Parameters included daily precipitation (P, mm), daily average temperature (Tmean, °C), daily average relative humidity (RH, %), daily average atmospheric pressure (AP, hPa), daily average wind speed (WS, m/s), and daily total solar radiation (GHI, MJ/m²).
  • Dynamic Step Size Adjustment (Dsa): Triggered by a daily rainfall threshold of 0.2 mm.

Main Results

• Feature Importance: SHAP analysis identified RTVICore (15.1 %), RH (11.9 %), and GCVI (10.4 %) as the top three most influential features for RSWC prediction.
• Optimal Input Window: An input window length of 10 days was determined as optimal for model performance.
• Model Performance (RSWC Inversion): The Transformer-DsaGRU model achieved the highest performance with a coefficient of determination (R²) of 0.81, a mean absolute error (MAE) of 1.12 % v/v, and a root mean square error (RMSE) of 1.65 % v/v, outperforming Transformer-GRU (R²=0.77), GRU (R²=0.69), LSTM (R²=0.68), RNN (R²=0.63), and TCN (R²=0.44).
• Short-term Prediction Accuracy:
  • For 1-day ahead forecasts, Transformer-DsaGRU achieved R²=0.78 and MAE=1.11 % v/v.
  • For 2-day ahead forecasts, Transformer-DsaGRU achieved R²=0.75 and MAE=1.06 % v/v.
  • This represents a 23.8 % increase in R² for 1-day ahead prediction compared to traditional GRU (0.63) and a 13.3 % reduction in MAE compared to Transformer-GRU (1.28).
• Generalization Ability: The model demonstrated robust generalization to an unseen year (2025 data), maintaining acceptable skill across different phenological stages, with an R² of 0.53 during the young fruit development stage.
• Computational Efficiency: The Dsa mechanism reduced the relative training time of Transformer-DsaGRU by 6.2 % compared to Transformer-GRU, optimizing computational resources.

Contributions

• Developed an efficient and interpretable feature engineering process using Pearson correlation, mutual information, and SHAP analysis to select 12 core features for RSWC prediction.
• Introduced a novel rainfall-threshold-based (0.2 mm) Dynamic Step Size Adjustment (Dsa) mechanism that adaptively extends the input sequence length from 10 to 12 days during rainfall events, effectively capturing the 0–3 day infiltration lag.
• Designed a cross-modal Transformer-GRU hybrid architecture that synergistically fuses UAV multispectral and meteorological time series, enabling the capture of both global spectral-meteorological relationships and short-term moisture dynamics.
• Achieved high-precision 1-2 day ahead RSWC forecasts at the plant scale, providing an approximately 3-day scheduling window for drip irrigation during the kiwifruit fruit expansion period.
• Demonstrated the feasibility and transfer potential of the "multi-source data + deep learning + dynamic step-size scheduling" framework for precision water and fertilizer management in kiwifruit orchards.

Funding

• Key Research and Development Projects of Shaanxi Province (Grant Nos. 2024NC-ZDCYL-05-03 and 2025PT-ZCK-25)
• Basic Research and Applied Basic Research Project of Hohhot City (Grant No. 2024-Guiji-34)
• National Center of Pratacultural Technology Innovation (under preparation)
• Special fund for innovation platform construction (CCPTZX2024N01)
• Key Programs of the Joint Fund of the National Natural Science Foundation of China (Grant No. U2243235)

Citation

@article{Li2026Research,
  author = {Li, X M and Chen, Zili and He, Jingyuan and Liu, Qingyuan and Niu, Zhen and Niu, Zhen and Gao, Zhilong and Jia, Zefeng and niu, zijie and niu, zijie and Zhang, Dongyan and Zhou, Mingu},
  title = {Research on inversion and prediction of root region soil water content in kiwifruit based on hyperparameter tuning by transformer-DsaGRU},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2025.110080},
  url = {https://doi.org/10.1016/j.agwat.2025.110080}
}