Sorghum yield prediction using UAV multispectral imaging and stacking ensemble learning in arid regions.

Area of Science:

• Agronomy
• Remote Sensing
• Machine Learning

Background:

• Drought and climate fluctuations challenge sorghum yield stability.
• Accurate, spatially explicit yield predictions are crucial for precision agriculture and food security.

Purpose of the Study:

• To develop a "spectral-meteorological-spatial" framework for sorghum yield prediction.
• To assess the effectiveness of machine learning algorithms and identify optimal monitoring stages.
• To provide technical support for precision sorghum management in arid environments.

Main Methods:

• Utilized multispectral imagery from a DJI Mavic 3M UAV and meteorological data.
• Collected data during key sorghum growth stages: seedling emergence, jointing, flowering, and maturity.
• Developed a 3D prediction framework using eight machine learning algorithms, employing SHAP values for variable importance analysis.

Main Results:

• Ensemble learning models, particularly Gradient Boosting (R² = 0.9491), showed superior performance.
• DVI and NDGI spectral indices were key predictors, with the jointing stage yielding the highest accuracy (R² = 0.9454).
• Yield predictions ranged from 4,291 to 4,965 kg ha⁻¹, exhibiting moderate positive spatial autocorrelation.

Conclusions:

• Integrating UAV multispectral data with machine learning offers efficient sorghum yield prediction.
• The jointing growth stage is optimal for monitoring and prediction accuracy.
• This approach supports precision planting and efficient sorghum management in arid regions.