Wang et al. (2025) Machine learning-based inversion and sensitivity analysis of soil moisture in Hemerocallis cultivation

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-11-27
• Authors: Jingshu Wang, Peng He, Keming Du, Xuran Li, Lishuai Xu, Fan Yang, Rutian Bi
• DOI: 10.1016/j.agwat.2025.110011

Research Groups

• College of Resources and Environment, Shanxi Agricultural University, Taigu, Jinzhong, China
• Institute of Environment and Sustainable Development in Agriculture, Chinese Academy of Agricultural Sciences (IEDA,CAAS), Beijing, China
• Rural Energy and Environment Agency, Ministry of Agriculture and Rural Affairs, Beijing, China
• Institute of Desert Meteorology, China Meteorological Administration/National Observation and Research Station of Desert Meteorology, Taklimakan Desert of Xinjiang/Taklimakan Desert Meteorology Field Experiment Station of China Meteorological Administration/Xinjiang Key Laboratory of Desert Meteorology and Sandstorm/Key Laboratory of Tree-ring Physical and Chemical Research, China Meteorological Administration, Urumqi, China

Short Summary

This study developed a machine learning framework integrating TSM640 sensor data, multi-source remote sensing (Sentinel-1/2), and meteorological datasets to analyze soil moisture (SM) dynamics across soil layers and their impact on Hemerocallis yield. The framework successfully estimated SM (BPNN R² = 0.64) and predicted yield (RF R² = 0.63), identifying bolting and squaring as critical moisture-sensitive growth stages for optimizing irrigation.

Objective

• To develop a machine learning framework for accurate monitoring and prediction of soil moisture (SM) dynamics across soil layers and their impact on yield in Hemerocallis cultivation.
• To identify water-sensitive growth stages for Hemerocallis to optimize irrigation strategies and improve agricultural productivity.

Study Configuration

• Spatial Scale: Hemerocallis planting areas (approximately 11 km²) in Yunzhou District, Datong City, Shanxi Province, China. Five sampling sites were selected, with soil moisture monitored at five depths: 0–20 cm, 20–40 cm, 40–60 cm, 60–80 cm, and 80–100 cm.
• Temporal Scale: Hemerocallis growing season in 2023. Meteorological data from May to August 2023. Sentinel-1 and Sentinel-2 satellite imagery for the same period. Soil moisture data collected every 10 minutes. Field yield survey conducted from July 19 to July 21, 2023.

Methodology and Data

• Models used: Backpropagation Neural Network (BPNN), Random Forest (RF), Support Vector Machine (SVM).
• Data sources:
  • Ground-based Soil Moisture: TSM640 tubular SM sensors and G30P data collector (Beijing Shiyutong Technology Co., Ltd., Beijing). Soil ring knife measurements for calibration.
  • Remote Sensing:
    • Sentinel-1 Interferometric Wide (IW) mode Ground Range Detected (GRD) products (VV, VH backscatter coefficients).
    • Sentinel-2 Level 2A (L2A) products (B2, B3, B4, B8 reflectance, and 15 derived spectral indices).
  • Meteorological Data: Daily precipitation (P), average temperature (Tavg), average pressure (Pavg), average relative humidity (RHavg), wind speed (WS), and sunshine duration (SD) from "China Surface Climate Daily Datasets" (116 weather stations). Interpolated using ANUSPLIN.
  • Elevation Data: Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM).
  • Yield Data: Field survey from 34 measured samples of Hemerocallis.

Main Results

• Soil Moisture Dynamics: The 0–60 cm soil layer exhibited the most significant soil moisture (SM) fluctuations, with a rate of change (f’(SM)) ranging from -3 %/d to 3 %/d, indicating it as the primary water-activity zone. SM dynamics were predominantly influenced by external irrigation management rather than planting duration.
• SM Modeling Performance: The Backpropagation Neural Network (BPNN) model demonstrated robust performance for SM estimation. The Sentinel-BPNN model for the 0–60 cm soil layer achieved a test set R² = 0.64, MAE = 1.89 %, MBE = 0.91 %, and RMSE = 2.60 %. Integrating multi-source data (Sentinel-1 and Sentinel-2) significantly improved accuracy, increasing R² by 0.15 and reducing error metrics (MAE: -0.57 %, MBE: -0.81 %, RMSE: -0.71 %) compared to single-source approaches. Meteorological data alone showed limited predictive capability (R² < 0.33) and minimal enhancement when combined with remote sensing data.
• Yield Prediction Performance: The Random Forest (RF) algorithm outperformed BPNN and SVM for Hemerocallis yield prediction, achieving a test set R² = 0.63, MAE = 0.0299 kg/m², MBE = 0.0078 kg/m², and RMSE = 0.0367 kg/m².
• Sensitivity Analysis of SM on Yield: Feature importance analysis identified the bolting stage (0.42) and squaring stage (0.32) as the most critical moisture-sensitive periods for Hemerocallis yield, followed by the filling stage (0.28) and seedling stage (0.25). Targeted irrigation during bolting and squaring is crucial for enhancing productivity.

Contributions

• Developed a novel machine learning framework for comprehensive soil moisture inversion across multiple soil depths (0–100 cm, with a focus on the 0–60 cm active zone) and yield prediction in Hemerocallis cultivation.
• Demonstrated the superior accuracy of integrating multi-source remote sensing data (Sentinel-1 radar and Sentinel-2 optical) for soil moisture estimation compared to single-source data or meteorological data alone.
• Quantified the impact of soil moisture at different growth stages on Hemerocallis yield, specifically identifying bolting and squaring as critical water-sensitive periods, providing actionable insights for precision irrigation.
• Bridged existing knowledge gaps in remote sensing-based soil moisture inversion for deeper soil layers and crop-specific water demand analysis, contributing to sustainable agricultural development.

Funding

• Innovation Project of Major State Basic Research Development Program (2021YFD1600301)
• Shanxi Basic Research Program (202403021212308)
• Research Topics of Shanxi Provincial Association for Science and Technology (KXKT202418)
• Science and Technology Innovation Fund of Shanxi Agricultural University (2024BQ74)
• 2024 Provincial Doctoral Innovation Project (2024KY301)
• National Natural Science Foundation of China (U2242209, 42301465)

Citation

@article{Wang2025Machine,
  author = {Wang, Jingshu and He, Peng and Du, Keming and Li, Xuran and Xu, Lishuai and Yang, Fan and Bi, Rutian},
  title = {Machine learning-based inversion and sensitivity analysis of soil moisture in Hemerocallis cultivation},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110011},
  url = {https://doi.org/10.1016/j.agwat.2025.110011}
}