Accurately predict the shelf-life of apple is urgently needed in practice. A feasible and non-equipment-depended data collection and model construction method was explored for shelf-life prediction of apple based on quality attributes observations and storage temperature. ‘Fuji’ apples were stored at four different temperatures of 0℃, 5℃, 15℃ and 25℃, respectively. The firmness, soluble solids content, titratable acid, SSC-TA rate, reducing ascorbic acid, starch content, weight loss and color values (L, a, b, ΔE, C) were measured periodically to obtain data set of 12 quality features at each storage stage and temperature. Feature selection method of SPCA and ReliefF was used to rank the quality attributes, respectively. Generative adversarial networks (GAN)-back propagation artificial neural network (BP-ANN), and BP-ANN were used to construct regression models between quality feature, storage temperature and shelf-life. Ratio of training set to test set was 3∶1. Totally 12 quality attributes were ranked in different orders by different feature selection methods. Using each accumulative combination of 1~12th quality attributes and storage temperature as input variables of GAN-BP-ANN and BP-ANN respectively, error rate of the validation set as evaluation criterion of prediction model. The accuracy of the models constructed by feature selection methods of SPCA were higher than that of ReliefF. The accuracy of the models established by GAN-BP-ANN were generally higher than that of the BP-ANN. It showed that GAN can effectively reduce the overfitting of BP-ANN model. Using three selected feature combinations as input variables, respectively, BP-ANN reached an accuracy above 0.930. GAN added BP-ANN can be a novel approach for accurately predict the shelf-life of postharvest “Fuji” apples by using the selected quality attributes and temperature.