To address the mechanistic limitations and data redundancy issues in the quantitative retrieval of soil Cd using hyperspectral remote sensing, an inversion method was proposed based on organic matter characteristic spectral bands. The method involved the extraction of characteristic spectral bands of organic matter with adsorption effects on heavy metal Cd in soil spectra. Subsequently, competitive adaptive reweighted sampling (CARS) was employed to optimize the selected spectral bands, and a partial least squares regression (PLSR) model was developed for the inversion of heavy metal Cd. The proposed method was validated by using laboratory spectral data from the Chenzhou mine and field spectral data from the Hami Huangshan South mine. The results demonstrated that the extraction of organic matter characteristic spectral bands not only reduced data redundancy but also significantly improved the accuracy of Cd inversion. In comparison to the correlation coefficient (CC) and genetic algorithm (GA) methods, the CARS algorithm exhibited superior performance in feature selection and inversion accuracy. The validation accuracies, expressed as R2, were 0.94 for the Chenzhou laboratory spectral data and 0.80 for the Hami field spectral data, indicating the robustness of the CARS-PLSR algorithm for both laboratory and field spectra. The findings can provide valuable references for feature band selection and algorithm optimization in the hyperspectral estimation of soil heavy metal content. The proposed method effectively addressed the limitations of existing approaches by leveraging the unique spectral characteristics of organic matter in soil.