Tunca et al. (2025) Integration of UAV images and ensemble learning for root zone soil moisture estimation in sorghum

Identification

• Journal: Irrigation Science
• Year: 2025
• Date: 2025-11-28
• Authors: Emre Tunca, Eyüp Selim Köksal, Sakine Çetin Taner
• DOI: 10.1007/s00271-025-01052-7

Research Groups

• Biosystem Engineering, Agriculture Faculty, Düzce University, Düzce, Turkey
• Department of Agricultural Structures and Irrigation, Agriculture Faculty, Ondokuz Mayıs University, Samsun, Turkey

Short Summary

This study developed and evaluated a methodology to estimate root-zone soil moisture in sorghum using high-resolution unmanned aerial vehicle (UAV) multispectral and thermal imagery combined with machine learning. An ensemble model integrating XGBoost, Light Gradient Boosting Machine, and K-Nearest Neighbors achieved the highest accuracy (R² = 0.85, RMSE = 11.124 mm/90 cm, MAE = 8.775 mm/90 cm) for field-scale monitoring.

Objective

• Identify optimal input features using exhaustive feature selection.
• Optimize and compare the performance of individual machine learning algorithms.
• Construct ensemble models and validate their performance against ground measurements for root-zone soil moisture estimation in sorghum.

Study Configuration

• Spatial Scale: Field-scale experiment with 6.3 m × 10 m (63 m²) plots; UAV imagery acquired at 40 m above ground level (AGL) providing high spatial resolution; root-zone soil moisture monitored at 0–90 cm depth.
• Temporal Scale: Two-season field experiment conducted during the sorghum growing periods of 2020 and 2021 (May to September); 21 synchronized UAV missions and ground measurements (12 in 2020, 9 in 2021).

Methodology and Data

• Models used:
  • Individual Machine Learning (ML) algorithms: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGB), K-Nearest Neighbors (KNN).
  • Ensemble learning: Stacking technique combining the individual ML models.
• Data sources:
  • UAV multispectral imagery (Blue 475 nm, Green 560 nm, Red 668 nm, Red-edge 717 nm, NIR 840 nm) from MicaSense Altum sensor.
  • UAV thermal imagery (11 µm) from MicaSense Altum sensor.
  • Ground-based volumetric soil moisture measurements (0–120 cm in 30 cm increments, with 0–90 cm as operational root zone) using a neutron moisture meter.
  • Derived features: 14 Vegetation Indices (VIs), Crop Height Maps (CHM), Soil Temperature (Ts), Canopy Temperature (Tc), and Reference Evapotranspiration (ET₀) calculated using FAO-56 procedures.

Main Results

• Exhaustive feature selection identified 11 optimal input variables: Green, Red, NIR, Ts, ET₀, Crop Height, OSAVI, RVI2, TVI, GI, and DVI.
• After hyperparameter tuning, the Random Forest (RF) model achieved the best performance among individual models with R² = 0.84, RMSE = 11.22 mm/90 cm, and MAE = 9.32 mm/90 cm on the test set.
• The ensemble model combining XGBoost, Light Gradient Boosting Machine (LGB), and K-Nearest Neighbors (K-NN) (Ensemble Model-4) yielded the highest accuracy with R² = 0.853, RMSE = 11.124 mm/90 cm, and MAE = 8.775 mm/90 cm on the test set.
• Ensemble learning generally outperformed individual ML algorithms, demonstrating improved model performance and generalizability.
• Estimated soil moisture maps showed clear differences aligned with irrigation treatments (full irrigation plots had highest soil moisture, rainfed plots lowest), and responded to rainfall events.

Contributions

• Developed a robust and practical UAV and machine learning-based workflow for high-resolution root-zone soil moisture estimation in sorghum.
• Demonstrated that ensemble learning techniques can modestly enhance the accuracy and robustness of soil moisture estimation compared to standalone machine learning models.
• Identified an optimal set of UAV-derived multispectral, thermal, and structural features, combined with reference evapotranspiration, for accurate root-zone soil moisture prediction.
• Provided a valuable tool for field-scale soil moisture monitoring, with direct applications in optimizing irrigation scheduling and improving agricultural water management practices.

Funding

• Scientific and Technological Research Council of Turkey (TUBITAK) under Grant Number 118O831.

Citation

@article{Tunca2025Integration,
  author = {Tunca, Emre and Köksal, Eyüp Selim and Taner, Sakine Çetin},
  title = {Integration of UAV images and ensemble learning for root zone soil moisture estimation in sorghum},
  journal = {Irrigation Science},
  year = {2025},
  doi = {10.1007/s00271-025-01052-7},
  url = {https://doi.org/10.1007/s00271-025-01052-7}
}

Generated by BiblioAssistant using gemini-2.5-flash (Google API)