Li et al. (2025) Estimating soil water content of cotton fields using UAV-based multi-source remote sensing data fusion

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-11-20
• Authors: Zhenxiao Li, Qian Cheng, Zhen Chen, Youzhen Xiang, Xiaotao Hu, Naftali Lazarovitch, Jingbo Zhen
• DOI: 10.1016/j.agwat.2025.109996

Research Groups

• Key Laboratory of Agricultural Soil and Water Engineering in Arid and Semiarid Areas of the Ministry of Education, Northwest A&F University, China
• Key Laboratory of Water-Saving Agriculture of Henan Province, Key Lab of Water-saving Irrigation Engineering, Ministry of Agriculture & Rural Affairs, Institute of Farmland Irrigation, Chinese Academy of Agricultural Sciences, China
• French Associates Institute for Agriculture and Biotechnology of Dryland, Jacob Blaustein Institutes for Desert Research, Ben-Gurion University of the Negev, Israel

Short Summary

This study aimed to improve soil water content (SWC) estimation in cotton fields by fusing UAV-based multi-source remote sensing and meteorological data with machine learning. It found that the CatBoost model, integrating multidimensional indices, achieved superior SWC estimation accuracy (R² = 0.762 ± 0.026) and robustness across different growth stages and irrigation levels.

Objective

• To systematically evaluate machine learning algorithms for improving soil water content (SWC) estimation accuracy by multidimensionally fusing thermal infrared, multispectral, and meteorological data.
• To calculate generalized normalized difference vegetation indices (NDVIi, j) from multispectral band combinations to identify optimal feature sets correlating with SWC.
• To develop a three-dimensional drought index (TDDI) based on NDVI, crop water stress index (CWSI), and air temperature (Ta) to analyze its dynamic response characteristics across soil layers and screen sensitive parameters.
• To compare the prediction accuracy of multiple machine learning models at different soil depths and evaluate model stability under varying conditions.

Study Configuration

• Spatial Scale: Field trials conducted over two years (2024-2025) at the Tianye Group Modern Water-Saving Agricultural Technology Demonstration Zone (85°59′14″ E, 44°19′55″ N), Xinjiang Uygur Autonomous Region, China. Experimental plots measured 13.4 m × 6.9 m in 2024 and 5 m × 6.9 m in 2025. UAV imagery provided centimeter-scale spatial resolution.
• Temporal Scale: Two-year field trial (2024-2025) covering the entire cotton growth cycle (approximately 150 days), with specific focus on critical cotton growth stages: flowering, boll setting, and boll opening. UAV image acquisition was conducted between 12:00 and 14:00 local time on specific dates during these stages. Soil water content was measured at multiple points and depths during these critical stages.

Methodology and Data

• Models used:
  • Support Vector Regression (SVR)
  • Random Forest Regression (RFR)
  • Gradient Boosting Decision Tree (GBDT)
  • Categorical Boosting (CatBoost)
• Data sources:
  • UAV Remote Sensing:
    • DJI Mavic 3 M: Multispectral imagery (Green: 560 nm, Red: 650 nm, Red edge: 730 nm, Near infrared: 860 nm).
    • DJI Mavic 3 T: Thermal infrared imagery (12 µm band).
  • Field Observations:
    • Soil Water Content (SWC): Gravimetric analysis of soil samples collected at six depths (0–10 cm, 10–20 cm, 20–30 cm, 30–40 cm, 40–50 cm, and 50–60 cm).
    • Meteorological data: Daily air temperature, wind speed, saturation vapor pressure, actual vapor pressure, net radiation flux, and soil heat flux density from an automatic weather station.
  • Derived Indices:
    • Normalized Difference Vegetation Index (NDVI) and generalized NDVI (NDVIi, j) from multispectral bands.
    • Crop Water Stress Index (CWSI) from thermal infrared imagery and environmental conditions.
    • Temperature Vegetation Dryness Index (TVDI) from vegetation indices and CWSI.
    • Three-Dimensional Drought Index (TDDI) from air temperature, NDVI, and CWSI.

Main Results

• The TVDIG, R index showed the strongest correlation with soil water content (SWC) (r = -0.47 ± 0.03), with heightened sensitivity at 0–10 cm soil depth (r = -0.5).
• The CatBoost model demonstrated superior SWC estimation performance (R² = 0.762 ± 0.026), significantly outperforming SVR (R² = 0.668 ± 0.053), RFR (R² = 0.659 ± 0.041), and GBDT (R² = 0.636 ± 0.033).
• CatBoost also achieved the lowest Mean Absolute Error (MAE = 0.757 ± 0.065) and Root Mean Square Error (RMSE = 1.050 ± 0.082), indicating superior robustness.
• The model's prediction consistency (Pc) was highest under low irrigation (I1, Pc = 0.915 ± 0.035) compared to medium (I2, Pc = 0.819 ± 0.030) and high (I3, Pc = 0.756 ± 0.063) irrigation levels.
• Prediction consistency improved significantly as the cotton growth cycle advanced, with Pc increasing from the flowering stage (0.678 ± 0.088) to the boll opening stage (0.946 ± 0.016).
• Deep soil layer SWC increased from 15.5 % under low irrigation (I1) to 18.2 % under high irrigation (I3) and from 16.3 % under high nitrogen (N3) to 17.2 % under low nitrogen (N1).
• Overall SWC decreased progressively from the flowering stage (15.6 %) to the boll opening stage (14.0 %).

Contributions

• Developed a synergistic inversion model for multi-depth SWC estimation by integrating multi-source remote sensing data (thermal infrared, multispectral) and meteorological data with machine learning algorithms.
• Explored the application potential of remote sensing features constructed through multi-band linear combinations by establishing a generalized NDVI (NDVIi, j), demonstrating that green-red spectral bands showed superior correlation with SWC compared to conventional indices.
• Showcased that the integration of multidimensional indices (TVDI, TDDI) effectively addressed the limitations of one-dimensional indices in retrieving remote sensing information for deeper soil layers.
• Identified the CatBoost ensemble learning model as highly effective for high-precision SWC estimation, particularly for SWC20, by resolving complex nonlinear interactions among spectral indices, canopy temperature, and meteorological data.
• Confirmed the robustness and practical applicability of the developed framework for precision irrigation decision-making in arid-region cotton fields across various growth stages and irrigation levels.

Funding

• Intelligent Irrigation Water and Fertilizer Digital Decision System and Regulation Equipment (2022YFD1900404)
• Postdoctoral Science Foundation Project of Shaanxi Province (2024BSHSDZZ230)
• National Key R&D Program of China (2023YFD1900705)
• Science and Technology Research Project of Henan Province (242102110355)
• Open Project of Henan Provincial Key Laboratory of Water-saving Agriculture (LWSAHN-2024–02)
• Natural Science Foundation Basic Research Project of Shaanxi Province (2023-JC-QN-0377)
• Planning and Budgeting Committee of the Council for Higher Education in Israel (VATAT) as part of the program ‘Israeli Center for Digital Agriculture’

Citation

@article{Li2025Estimating,
  author = {Li, Zhenxiao and Cheng, Qian and Chen, Zhen and Xiang, Youzhen and Hu, Xiaotao and Lazarovitch, Naftali and Zhen, Jingbo},
  title = {Estimating soil water content of cotton fields using UAV-based multi-source remote sensing data fusion},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.109996},
  url = {https://doi.org/10.1016/j.agwat.2025.109996}
}