Wheat growth parameters prediction based on dual output Bayesian neural network using multi-modal information.

Area of Science:

• Agricultural Science
• Remote Sensing
• Plant Physiology

Background:

• Leaf Area Index (LAI) and Leaf Nitrogen Accumulation (LNA) are critical for assessing wheat growth and nitrogen status.
• Existing prediction models often use single data sources, limiting comprehensive crop trait characterization.

Purpose of the Study:

• To develop a multimodal learning framework for simultaneous and accurate prediction of wheat LAI and LNA.
• To enhance crop growth and nitrogen status monitoring through integrated data analysis.

Main Methods:

• Extracted spectral, image, and canopy structural features from wheat.
• Developed a canopy height correction method for structural feature extraction.
• Constructed a Dual-Output Bayesian Neural Network (DO-BNN) with Extreme Sample Mining (ESM) and a joint loss function.

Main Results:

• Fused multimodal data in the DO-BNN yielded high predictive performance (R² of 0.89 for LAI, 0.77 for LNA).
• Achieved lower errors (RRMSE of 0.15 for LAI, 0.35 for LNA) compared to single-modal methods.
• Demonstrated superior accuracy and robustness in predicting wheat growth parameters.

Conclusions:

• Integrating spectral, image, and structural data improves wheat canopy trait characterization.
• The DO-BNN effectively models the relationship between crop growth and nitrogen accumulation.
• The framework offers a promising approach for high-accuracy, collaborative wheat monitoring.