Hyperspectral imaging technology has emerged as an non-destructive and reliable analysis and discriminant technology for agri-food safety assessment. The technology provides the 3D cub e data with spatial and spectral data. The size of the 3D cube data set is larger than hundreds MB (in the case of 1000 pixels × 1004 pixels × 128 bands). The real-time detection and classification technology is essential for food safety assessment. To reduce the size of data is finding the optimal bands from the whole spectral data. In this study, a genetic algorithm (GA) is implemented to find subset features from 128 wavelengths and applied to develop a classification model. On the stainless steel plate, six spinach droplets were placed on each well according to the concentrations. Original juice of spinach is 100%, and additional five levels were diluted with distilled water as follows: 1:5 (20%), 1:10 (10%), 1:20 (5%), 1:50 (2%), and 1:100 (1%), respectively. VNIR hyperspectral images were obtained using a line-scan hyperspectral imaging system and concentration prediction models were developed with multivariate analysis methods. Support vector machine with 39 selected bands using the genetic algorithm showed accuracy a s 90.65% and the kappa coefficient was 0.88. The overall accuracy of PLS-DA and LDA showed reasonable accuracy as 72.13% and 85.06%, respectively. Using feature selection such as gen etic algorithm, we can reduce the dimensionality of the 3D cube data so that it is helpful to develop a rapid and real-time classifier for food safety. VNIR (400-1000 nm) hyperspectral imaging system and chemometric classification models with sub-set data based on genetic algorithm showed a potential for developing an safety assessment technology for agro-food processing machines or facilities.