Bayesian modelling of phosphorus content in wheat grain using hyperspectral reflectance data.

Area of Science:

• Agricultural Science
• Remote Sensing
• Statistical Modeling

Background:

• Technological advancements have increased sensor use in crop surveys for agricultural data modeling.
• Plant spectroscopy combined with statistical modeling offers a cost-effective alternative to direct measurement for assessing plant chemical components.
• This research focuses on modeling phosphorus content in wheat grain using hyperspectral sensor reflectance data.

Purpose of the Study:

• To model phosphorus content in wheat grain using hyperspectral reflectance data.
• To identify the most important spectral bands for phosphorus prediction using a Bayesian variable selection procedure.
• To compare the predictive performance of different spatial correlation models against independent observations.

Main Methods:

• Utilized hyperspectral sensor data to measure wheat grain reflectance at various wavelengths.
• Employed a Bayesian variable selection procedure to identify key spectral bands.
• Evaluated three statistical models: independent observations, Conditionally Autoregressive (CAR) model, and an exponential correlogram model for spatial dependence.
• Assessed model goodness-of-fit using the Deviance Information Criterion (DIC) and predictive power via cross-validation.

Main Results:

• The Conditionally Autoregressive (CAR) model demonstrated the best fit and predictive performance for phosphorus content.
• Variable selection within the CAR model identified critical wavelengths between 500-690 nm.
• The CAR model's predictive accuracy surpassed traditional vegetative indices by an average of 14.7% across three years (2010-2012).

Conclusions:

• The proposed Bayesian approach for predicting phosphorus content in wheat grain is a viable and effective alternative.
• The CAR model, incorporating spatial dependence, significantly enhances prediction accuracy compared to models assuming independence or standard vegetative indices.