Identification of multiple defects in Nanfeng mandarin based on fusion of spectral data and texture data

The external quality of fruits is a crucial criterion for consumer purchasing power and reflects their market value. This study proposes a method for identifying defects in Nanfeng mandarins by integrating hyperspectral technology and data fusion. Four preprocessing techniques—Savitzky-Golay smoothing (SGS), standard normal variate (SNV), orthogonal signal correction (OSC), and Savitzky-Golay first derivative (SGFD)—were applied to spectral data. Dimensionality reduction was performed using four distinct methods. Additionally, texture features were extracted from the first two principal component images of Nanfeng mandarins using a gray-level co-occurrence matrix (GLCM). Classification models for defect recognition were developed using partial least squares discriminant analysis (PLS-DA), random forest (RF), and convolutional neural network (CNN) based on spectral data, texture data, and fused data (spectral + texture). The results demonstrated that classification models relying on spectral data outperformed those using texture data. Notably, the CNN model established with fused data proved most effective in identifying defects in Nanfeng mandarins, achieving an overall accuracy of 98 %, with a classification accuracy exceeding 95 % for all four defect types. The combination of spectral features and texture features significantly enhances the discriminative performance of the model. • Applied multiple preprocessing and feature selection methods to spectral data. • Extracted 14 texture features from images using gray-level co-occurrence matrix (GLCM). • Developed PLS-DA, RF, and CNN models based on spectral and texture data separately. • The CNN model established with fused data achieved most effective, achieving an overall accuracy of 98 %.