2026 Volume 7 Issue 3 Pages 63-74
For optimization of processing applications of soybeans, techniques that can rapidly and non-destructively evaluate composi-tional properties are required. In this study, the aim was to construct estimation models of soybean component amounts using fluorescence imaging with multiple excitation wavelengths and machine learning. One hundred soybean samples from Hokkaido were used, and excitation wavelengths of 365 nm and 420 nm were selected based on excitation fluorescence matrices. Color and fluorescence images obtained were used as inputs, and estimation models using partial least squares regression and convolu-tional neural networks were constructed. As a result, PLSR showed superior performance, and 𝑅2 = 0.66 and RMSE = 0.78% for lipid content, and 𝑅2 = 0.47 and RMSE = 24.09 mg/100 g FW for isoflavone content were obtained, and it was shown that information contained in fluorescence images may be effective for estimation of soybean component amounts.
