Textural firmness is a primary determinant of consumer acceptance for evaluating freshness quality of fish fillet flesh. The objective of this study was to investigate the potential of using visible and near-infrared hyperspectral imaging (400–1000 nm) for non-destructive prediction of firmness quality of grass carp fillet as affected by frozen storage. Fillet samples were frozen at − 20 °C for 24 h and then stored at 4 °C for thawing over five days. Hyperspectral images were obtained at different thawing stages and their corresponding spectral data were extracted. Two calibration models were established between the extracted spectral data and the reference firmness values measured by the traditional mechanical method by using partial least squares regression (PLSR) and least-square support vector machine (LS-SVM) analysis. Three approaches of regression coefficients (RC) from PLSR analysis, genetic algorithm (GA) and successive projection algorithm (SPA) were utilized to recognize the most important wavelengths that possessed the greatest influence and sensitivity on the firmness prediction based upon the whole spectral range. By comparing the above-mentioned three variable selection methods, seven optimal wavelengths (450, 530, 550, 616, 720, 955 and 980 nm) were selected by GA and its corresponding simplified prediction model of GA-LS-SVM was also obtained, showing the best performance with the highest determination coefficient (R2P) of 0.941 and the lowest root mean square error estimated by prediction (RMSEP) of 1.229. The overall results of this study suggested that hyperspectral imaging technique has a potential for fast and non-destructive prediction and analysis of textural firmness of grass carp fillets as affected by frozen storage.