Abstract: In response to the problem that the detection of soluble solids content (SSC) in Gongli pears using visible/near-infrared spectroscopy (Vis/NIRS) is vulnerable to interferences from sample temperature fluctuations, a temperature correction method was proposed by integrating competitive adaptive reweighted sampling (CARS) with a one-dimensional convolutional neural network (1D-CNN) regression model, and six temperature gradients (5, 10, 15, 20, 25, and 30 ℃) were established for validation. The introduction of temperature labels as auxiliary variables in the model input helped the neural network perceive and adapt to spectral changes under different temperature conditions, thereby enhancing the robustness of the model to temperature perturbation. This method was compared and validated against other temperature correction methods such as global calibration, generalized least squares weighting (GLSW), and external parameter orthogonalization (EPO). The results showed that CARS-1D-CNN outperformed traditional methods such as EPO in terms of prediction accuracy and robustness, with correlation coefficient of prediction (Rp) of 0.885 9 and root mean square error (RMSE) of 0.548 3. Compared with the traditional method EPO used in this study, CARS-1D-CNN improved the correlation coefficient by 2.96% and reduced the prediction root mean square error of prediction by 2.73%. This method effectively mitigates the interference of temperature on the spectral model, improving its stability and prediction performance.

Key words: competitive adaptive reweighted sampling combined with 1D convolutional neural network; visible/near infrared spectroscopy; soluble solids content; temperature correction