Enhancing model robustness through different optimization methods and 1-D CNN to eliminate the variations in size and detection position for apple SSC determination

Abstract

The visible/near-infrared (Vis/NIR) spectroscopy technique has been widely used for the online detection of soluble solids content (SSC) in apples. However, external factors such as sample size and detection position can cause spectral distortion, resulting in a decline in detection accuracy. In this study, we aimed to develop a more robust prediction model that can resist the impact of sample size and detection position on the model. Firstly, we collected and analyzed the transmission spectra of apple samples under different sizes and detection positions using a self-designed Vis/NIR spectroscopy online acquisition device. It was found that the effect of fruit size and detection position on Vis/NIR spectra was due to optical path difference. Thus, a diameter correction method was utilized to uniformly correct the obtained absorbance spectra. The performance of local models achieved better results after correction. And then, global models with various preprocessing methods were developed. To further improve the model performance, changeable size moving window (CSMW) and competitive adaptive reweighted sampling (CARS) were utilized to select the effective wavelengths. After that, one dimensional-convolutional neural network (1D-CNN) model was constructed, which outperformed the other models without any preprocessing and optimization methods, and the values of RCal2, RMSEC, RPre2, and RMSEP are 0.953, 0.254 %, 0.900, and 0.371 %, respectively. In this study, conventional PLSR modelling methods and deep 1D-CNN method were compared under the influence of the two external factors, and the result showed that 1D-CNN can serve as a more convenient alternative for apple online SSC determination, which could significantly reduce the complexity of the Vis/NIR spectroscopy modeling process.