Abstract: To address the limitations of current methods for determining the physicochemical properties of cucumber fruits, which require sample destruction and are time-consuming, labor-intensive, and unsuitable for large-scale testing. This study included cucumber fruits grown under five irrigation gradients. By combining hyperspectral imaging (HSI) technology with chemometric methods, this study developed predictive models for the soluble solids content, soluble sugar content, and firmness of cucumber fruits. The results indicated that the optimal prediction models for soluble solids, soluble sugars, and firmness in cucumber fruits were established using a convolutional neural network (CNN). The coefficients of determination for the prediction set (Rp2) of the models were 0.861 0, 0.894 0, and 0.832 0, respectively, and the root mean square errors of prediction (RMSEP) were 0.145 0 °Brix, 0.001 0%, and 3.415 0 N, respectively. Furthermore, the models enabled the visualization of the spatial distribution of these indicators, revealing the spatial heterogeneity of cucumber quality under different irrigation treatments and verifying the excellent predictive capability and high reliability of the models. In conclusion, hyperspectral imaging combined with chemometrics can achieve rapid, non-destructive detection of cucumber fruit quality. This method is systematic and highly accurate, providing a technical basis for the precision irrigation management and quality improvement of vegetables.

Key words: cucumber; hyperspectral imaging technology; non-destructive testing; quality; convolutional neural network; deep learning