During the storage, under high temperature and hypoxia, potato internal flesh tends to become black. It seriously reduces the quality of processed potato products and the utilization of raw materials. Blackheart potatoes can not be distinguished from their appearance. The traditional detection method requires the potato to be cut to judge, which is only suitable for sampling inspection. Potato spectrum data were collected based on the self-developed potato internal quality spectrum detection device. The energy spectrum and absorbance spectrum data of 234 healthy potatoes and 236 blackheart potatoes were collected respectively. The sample set was divided into calibration set and validation set at a ratio of 3∶1 by random. The sensitivity, specificity, and classification accuracy were used as model evaluation indexes. Based on the absorbance spectrum, after pretreatment by Auto, the partial least squares-linear discriminant analysis (PLS-LDA) model for potato blackheart defect was established in the range of 500~950nm. The competitive adaptive reweighting sampling (CARS) algorithm and successive projection algorithm (SPA) were adopted jointly to screen key variables. As a result, the sensitivity, specificity, and total accuracy of the optimal discrimination model for blackheart potato, with 9 variables, reached 98.87%, 98.30% and 98.44%, respectively. Based on the energy spectrum, the dual-wavelength correlation analysis method was adopted. The energy difference and ratio of any wavelength pair were calculated for the correlation analysis of blackheart defect. Finally, the linear discriminant analysis (LDA) was established by the energy ratios of two variables T699/T435. The sensitivity, specificity and total accuracy of the discrimination model reached 97.71%, 96.15% and 97.67%, respectively. Therefore, both the CARS-SPA-PLS-LDA model based on the absorbance spectrum and the (T699/T435)-LDA model based on the energy spectrum could identify blackheart potato effectively. Compared with the absorbance spectrum model, the energy spectrum model used only two variables. It was simple, stable, and had a wide applicability, which solved the limits of the two reference, white and dark background.