Hyperspectral data are characterized by high dimensionality and richer feature information. This high-dimensional data provides more opportunities to improve classification accuracy and precision in vegetation classification. Traditional feature wavelength modelling often results in poor classification accuracy due to too many input variables. To overcome this problem and improve the ability of the model to capture the subtle spectral differences of wetland vegetation, this paper explores the east coast of the Erhai Lake as the study area and hyperspectral data of three typical wetland vegetation (Mizuno, Ruscus, and Sophora japonica) are measured as the target samples.The sample spectral curves were SG smoothing as original spectra (OS), continuum removal transform (CR), and first-order differentiation (FD) and analyzed for spectral features; then, the original spectra were decomposed by Variational mode decomposition (VMD) into 8 scales. Next, the wavelengths selected by the Competitive adaptive reweighted sampling (CARS) algorithm were used as the characteristic wavelengths. Finally, the best combination of parameters found was used to put into the Bayesian algorithm optimized support vector machine (Bayes-SVM) for modeling. The results show that the number of feature wavelengths extracted by the CARS algorithm is reduced, and most of them are distributed in the absorption feature intervals of vegetation, and the effect of dimensionality reduction is significant; the model constructed by the 4th mode after decomposition (S4-CARS-Bayes-SVM) has the best classification effect, with a precision rate (PR) of 0.933 3, a recall rate (RR) of 0.888 9, an F1 score of 0.896 3, and AUC value of 0.928 6, i.e., this model has strong robustness as well as recognition performance.