Rapid Detection of Deoxynivalenol in Soybeans by NIR Spectroscopy Coupled With ETs-PLS Feature Selection Method

This study developed a hybrid feature selection approach combining Extra Trees (ETs) with partial least squares (PLS) regression for rapid and accurate deoxynivalenol (DON) quantification in soybeans using near-infrared (NIR) spectroscopy. The ETs algorithm ranked spectral variables by importance, and the top 50 wavelengths were selected as inputs for the PLS model. The proposed ETs-PLS method was compared with three conventional wavelength selection methods, and the ETs-PLS achieved superior performance with R² _P of 0.950 and RMSE_P of 87.3 µg/kg, significantly outperforming full-spectra PLS and all benchmark methods. The selected wavelengths were concentrated in the 1139-1183 nm region, corresponding to C─H stretching vibrations. This ETs-PLS method offers an effective balance between model parsimony and predictive accuracy, demonstrating great potential for rapid and nonconsumptive mycotoxin screening in agricultural products.