Estimation of protein content in wheat samples using NIR hyperspectral imaging and 1D-CNN

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Wheat protein content is crucial for its market value and applications.
• Traditional protein analysis methods are time-consuming, require laboratory access, and are not real-time.
• Existing hyperspectral imaging (HSI) models for wheat protein estimation have limitations in protein range and geographic applicability.

Purpose of the Study:

• To extend the application of HSI for estimating wheat protein content over a wider range.
• To validate HSI models for wheat cultivated in diverse geographical regions.
• To develop and compare machine learning models for accurate, real-time protein estimation.

Main Methods:

• Acquired hyperspectral images of 621 wheat samples from five Indian regions (900-1700 nm).
• Determined reference protein content using the Kjeldahl method (range: 9.5-17.25%).
• Extracted mean spectra and developed/validated 1D-CNN and conventional machine learning models using 5-fold cross-validation.

Main Results:

• The 1D-CNN model demonstrated superior performance with R² of 0.9972, RMSE of 0.0771, and RPD of 18.81.
• The model accurately estimated protein content for wheat across a broad range and different cultivation regions.
• HSI combined with 1D-CNN provides a robust, non-destructive method for protein analysis.

Conclusions:

• One-dimensional convolutional neural networks (1D-CNN) effectively estimate wheat protein content using hyperspectral imaging.
• This HSI approach eliminates the need for wet labs and offers real-time analysis.
• The findings have significant implications for farmers, traders, and the food industry by enabling rapid quality assessment.