Retrogradation is an important physicochemical property of starch during processing, transportation and storage. Rapid detection of retrogradation is of great significance to starch products’ quality and shelf life. In order to investigate the feasibility of selecting the characteristic variables of retrograde starch by two-dimensional correlation spectroscopy (2D-COS), spectral fusion technology and 2D-COS was combined to quantitatively characterize the retrogradation characteristics of wheat starch in this study. First, wheat starch’s crystallinity and retrogradation degree at different retrograde times were measured. The retrograde properties of starch were characterized by crystal content in the starch system and resistance to amylase hydrolysis. Then, the samples’ near-infrared and mid-infrared spectral data were collected respectively. After spectral pretreatment, prediction models based on near-infrared, mid-infrared, -and fusion spectra were established using partial least squares analysis. On this basis, the retrogradation day was used as the external disturbance. Starch spectra of 0, 1, 2, 3, 5, 7, 10, 14, 21 and 35 days were selected for 2D-COS analysis. By analyzing the synchronization and autocorrelation spectrum, 13 and 11 feature variables related to starch retrogradation characteristics were identified from near-infrared and mid-infrared spectra, respectively. Finally, prediction models for retrogradation degree and crystallinity were established based on these variables. The results show that the models based on full-spectra yielded better prediction performance after spectral fusion, with relative percent deviation (RPD) increasing from 1.203 4 and 2.069 0 to 3.980 9 and from 2.594 0 and 2.109 9 to 4.576 3 for crystallinity and retrogradation degree. Using the feature spectra obtained by 2D-COS analysis, the RPD values for the crystallinity model and retrogradation degree model increased to 8.095 9 and 14.183 6. 2D-COS can improve spectral resolution and obtain more chemical structure information than the model based on Competitive Adaptive Reweighted Sampling. Therefore, the spectral fusion technology combined with 2D-COS model has better results. The results show that it is feasible to use the 2D-COS to identify the characteristic wavelengths for starch retrogradation properties, which provides a new idea for the characteristic variables optimization of fusion spectra. Spectral fusion technology combined with 2D-COS can realize the rapid detection of starch retrograde, which provides a method for rapidly detecting starch food quality and shelf life.