Knowledge-guided temperature correction method for soluble solids content detection of watermelon based on Vis/NIR spectroscopy

Visible/near-infrared (Vis/NIR) spectroscopy technology has been extensively utilized for the determination of soluble solids content (SSC) in fruits. Nonetheless, the spectral distortion resulting from temperature variations in the sample leads to a decrease in detection accuracy. To mitigate the influence of temperature fluctuations on the accuracy of SSC detection in fruits, using watermelon as an example, this study presents a knowledge-guided temperature correction method utilizing one-dimensional convolutional neural networks (1D-CNN). This method consists of two stages: the first stage involves utilizing 1D-CNN models and gradient-weighted class activation mapping (Grad-CAM) method to acquire gradient-weighted features correlating with temperature. The second stage involves mapping these features and integrating them with the original Vis/NIR spectrum, and then train and test the partial least squares (PLS) model. This knowledge-guided method can identify wavelength bands with high temperature correlation in the Vis/NIR spectra, offering valuable guidance for spectral data processing. The performance of the PLS model constructed using the 15 °C spectrum guided by this method is superior to that of the global model, and can reduce the root mean square error of the prediction set (RMSEP) to 0.324°Brix, which is 32.5 % lower than the RMSEP of the global model (0.480°Brix). The method proposed in this study has superior temperature correction effects than slope and bias correction, piecewise direct standardization, and external parameter orthogonalization correction methods. The results indicate that the knowledge-guided temperature correction method based on deep learning can significantly enhance the detection accuracy of SSC in watermelon, providing valuable reference for the development of PLS calibration methods.