The existing hyperspectral vegetation indices used for estimating the canopy leaf area index (LAI) of winter wheat (Triticum aestivum L.) performed well, but the use of such indices at late growth stages can lead to inaccurate results. To improve the performance of LAI models for wheat in late growth stages, the continuous wavelet transform (CWT) method was applied in this study and used to decompose the canopy reflectance and its first derivative into wavelet coefficients. The correlation scalograms of wavelet coefficients and the LAI were then constructed and used to extract the top 1% correlated region as the wavelet feature. The canopy LAI estimation model for late growth wheat was established at last and compared with models based on 12 different types of hyperspectral vegetation indices. The results showed that, compared with the estimation models using the hyperspectral vegetation indices (for which the values were all less than 0.15 and the root-mean-square errors (RMSEs) were greater than 1), the CWT-based canopy LAI estimation model for late growth wheat had obvious improvements in accuracy (maximum of 0.53 and minimum of RMSE of 0.78). Hence, this new method shows promise for use in agricultural and ecological applications.