To explore the effectiveness of spectral indices (SIs) combined with continuous wavelet transform (CWT) in soil organic carbon (SOC) inversion, difference indices (DI), ratio indices (RI), normalized difference indices (NDI), and re-normalized difference indices (RDI) are applied to process spectral data along with CWT for 61 soil samples from Deming Town in northeastern Germany. Through the analysis of two-dimensional correlation between these indices and wavelet coefficient and SOC content, the top 20 bands with the highest correlation in SIs and wavelet coefficients are selected. Models for SOC content inversion are constructed using input variables of original full spectrum (OR), SIs, CWT, and SIs+CWT, coupled with support vector machine (SVM), random forest (RF), and back propagation artificial neural network (BPANN) algorithms. Results indicate that after processing by SIs and CWT, spectral data have an increase in correlation coefficients by ~0.17 and ~0.14 compared to the original spectrum. Among the machine learning models, the BPANN exhibites the best model accuracy. Among the four input variables, the SIs+CWT-BPANN model demonstrates the best validation performance, whose coefficient of determination is 0.95, root mean square error is 2.13 g/kg, and relative analysis error is 4.67, indicating that the optimal feature coupling machine learning method combining SIs and CWT can improve the estimation accuracy of SOC content. This study offers a fresh perspective on accurately estimating SOC content, which holds critical significance in evaluating both soil quality and crop yield.