It is crucial to determine the content and distribution of heavy metals in soil in a timely and accurate manner. Based on GaoFen-5 hyperspectral images, this study investigates a large-scale inversion of soil Cd content in Tongguan County. The competitive adaptive reweighting algorithm and genetic algorithm (CARS-GA) are coupled via feature coding and random mutation to accurately screen the characteristic bands of Cd and improve the inversion accuracy of the model. The characteristic bands of Cd are searched for, first based on the global search strategy and then local search. Under the two spectral enhancement methods of standard normal variate (SNV) and first differential (FD), the accuracy values of the partial least squares (PLSR) models established using the CARS-GA method and other band selection methods (correlation coefficient analysis method and CARS algorithm) are compared. Finally, the optimal model is selected and applied to the entire bare-land area of Tongguan County. The experimental results show that when the CARS-GA method is used for band selection, the accuracy values of the PLSR model constructed based on the two spectral transformation datasets are significantly higher than those constructed using the correlation coefficient analysis method and the CARS algorithm. During the FD spectral transformation, the coefficient of determination of the validation set increases by 0.288 and 0.093, respectively. In the SNV-transformed spectrum, the coefficient of determination of the validation set increases by 0.372 and 0.088, respectively. This study demonstrates that the band selection based on the CARS-GA algorithm can effectively enhance the robustness of the Cd-content estimation model, providing improved data support for environmental pollution assessment and ecological protection.