A nonlinear classification method for online sorting of navel orange SSC based on near-infrared spectroscopy and data augmentation

As a non-invasive detection method, near-infrared spectroscopy (NIR) has demonstrated significant potential for application in assessing fruit quality and sorting. However, during the online fruit sorting process, multiple factors affect the sorting accuracy. To address the challenges of limited sample size, heterogeneous quality, and the intricate nonlinear relationship between detection indices and spectral data in online sorting, this paper presents a data augmentation approach for the online sorting of navel oranges based on the soluble solids content (SSC), which integrates error rate and probability mass weighting. Firstly, cluster analysis was performed on the SSC, and the R² statistic and the elbow rule were introduced to determine the optimal number of clusters. The training set and test set were divided using the Monte Carlo random sampling method. Subsequently, the training set was augmented by incorporating spectral data that had undergone moving average smoothing, thereby forming an enhanced sample set. Furthermore, the probability mass and error rate of training set samples were integrated to formulate the sample weight coefficient. At last, the augmented training set was employed to establish a classification model via a three-layer neural network. Multiple experimental results showed that this method significantly improves classification performance for online sorting data, and the classification accuracy exceeds 90 %.