Soil Moisture Inversion in Alfalfa via UAV with Feature Fusion and Ensemble Learning.

Area of Science:

• Agricultural Engineering
• Remote Sensing
• Machine Learning

Background:

• Precise irrigation relies on accurate soil moisture data.
• Unmanned aerial vehicle (UAV) remote sensing offers high spatiotemporal resolution for soil moisture monitoring.
• Alfalfa fields require tailored soil moisture management across various growth stages.

Purpose of the Study:

• To retrieve soil moisture content in alfalfa fields using UAV multispectral imagery.
• To evaluate the performance of machine learning and ensemble learning models for soil moisture inversion.
• To investigate the impact of integrating spectral and texture features on inversion accuracy.

Main Methods:

• Construction of a multi-source feature set by fusing spectral and texture data from UAV multispectral images.
• Systematic comparison of Random Forest Regression (RFR), K-nearest neighbors regression (KNN), and XG-Boost models.
• Evaluation of ensemble learning models (Voting and Stacking) for enhanced soil moisture retrieval.

Main Results:

• Ensemble learning models, particularly Voting, outperformed individual machine learning models, achieving a maximum R² of 0.874 and RMSE of 0.005.
• The fusion of multi-source features significantly improved inversion accuracy, increasing R² by 0.065 compared to single-texture features.
• Optimal soil moisture inversion depths varied by growth stage: 40-60 cm for branching and budding, 20-40 cm for early flowering.

Conclusions:

• Integrating multi-source feature fusion with ensemble learning provides a highly accurate and stable method for alfalfa soil moisture inversion.
• This approach offers an effective technical solution for precise water management in artificial grasslands within arid regions.
• The findings support the use of UAV-based remote sensing for optimizing irrigation in agricultural ecosystems.