Hybrid wavelength selection technique and spectral binning for wheat protein estimation using hyperspectral imaging

Hyperspectral imaging has shown potential for estimation of wheat protein content, but it requires expensive equipment and generates high-dimensional data. This study identifies a minimal set of informative wavelengths to reduce computational complexity and facilitate the development of low-cost spectral imaging systems. We employed thirteen wavelength selection algorithms and their combinations on the raw and preprocessed spectral data with a 5 nm resolution to identify optimal wavelengths. The best results were obtained with 6 wavelengths (R² = 0.9790, RMSE = 0.2104) using a two-step hybrid strategy combining Random Forest and Genetic algorithm coupled with support vector regression. The accuracy remained comparable (R² = 0.9688, RMSE = 0.2564) when the resolution was reduced to 10 nm using spectral binning. This indicates that six wavelengths and 10 nm resolution can be used for accurate estimation of the wheat protein content. These findings highlighted the potential for developing an inexpensive multispectral imaging device.