Multi-trait spectral modeling for estimating grapevine leaf traits and nutrients.

Area of Science:

• Plant physiology and remote sensing
• Agricultural science and precision farming
• Biophysical modeling and machine learning

Background:

• Leaf biochemical and mineral nutrient analysis is crucial for precise farm management.
• Hyperspectral imaging offers a non-destructive method for assessing plant health.
• Accurate estimation of grapevine leaf traits is vital for optimizing viticulture practices.

Purpose of the Study:

• To explore leaf spectral modeling for estimating biochemical and nutritional traits in grapevine leaves.
• To compare the predictive performance of single-trait versus multi-trait spectral modeling approaches.
• To develop and validate advanced machine learning models for trait prediction.

Main Methods:

• Collected hyperspectral data (400-2500 nm) from grapevine leaves across three growing seasons.
• Imputed missing trait data using the PROSPECT-PRO model and a Convolutional Neural Network (CNN).
• Reduced spectral bands and employed hybrid CNN-Long Short-Term Memory (LSTM) networks for single-trait and multi-trait predictions.

Main Results:

• The multi-trait model significantly outperformed single-trait models in predicting most grapevine leaf traits.
• Key traits like nitrogen, phosphorus, Nstruct, and manganese showed substantial improvements in prediction accuracy with the multi-trait approach.
• The multi-trait model leveraged shared spectral information and inter-trait dependencies for enhanced predictive power.

Conclusions:

• Multi-trait spectral modeling is a more efficient and accurate method for predicting grapevine leaf traits compared to single-trait models.
• This approach holds significant potential for advancing precision agriculture and farm management.
• Integrating radiative transfer models with advanced machine learning enhances the utility of hyperspectral data for plant analysis.