The accurate inversion of soil heavy metal pollution in hyperspectral analysis relies on carefully selecting characteristic band extraction methods and inversion models. Finding the optimal combination of these two factors to achieve the highest system inversion accuracy remains an urgent and essential problem in this field. The present study involved the collection of 92 sets of soil samples from a typical Chromium (Cr) contaminated area in South China. The Cr content was quantified using Inductively Coupled Plasma Mass Spectrometry (ICP-MS). Additionally, the ASD Field Spec4 Spectrometer was employed to gather hyperspectral information in the laboratory. The spectral information preprocessing employed the combined SG+SNV+SD method. Here, SG refers to the Savitzky-Golay smoothing filter, SNV stands for Standard Normal Variate normalization, and SD represents second-order derivative transformation. This combined methodology was employed on the unprocessed spectral data to diminish the impact of soil scattering and noise. Consequently, it enhanced both the quality of spectral data and the precision of feature analysis. Four algorithms, namely Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), Uninformative Variable Elimination (UVE), and Genetic Algorithm (GA) were employed to extract Characteristic bands. Subsequently, the relationships between the extracted Characteristic bands and Cr content were established by using four inversion models: Multivariate Linear Regression (MLR), Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Artificial Neural Network (ANN). A comparative analysis of various Characteristic band extraction methods and combinations of inversion models regarding their impact on the accuracy of soil Cr content inversion determined that the SG+SNV+SD preprocessing enhances the spectral data’s capability to represent characteristic information. CARS and UVE Characteristic band extraction methods can significantly enhance the predictive performance of PLSR, MLR, and SVR models. In contrast, the SPA method improves the predictive effectiveness of the ANN model. Through the combination approach of SG+SNV+SD+CARS+PLSR, a total of 98 characteristic bands located within the ranges of 800～1 000, 1 400～1 700, and 2 100～2 450 nm were extracted. Model validation yielded an R² value of 0.97, RMSE of 5.25 mg·kg^-1, MAE of 4.35 mg·kg^-1, and RPD of 3.94. These evaluation metrics demonstrate the exceptional predictive capability of the model for soil Chromium Cr. In this research, soil Cr pollution was selected as a case study for hyperspectral inversion. A comparative analysis of various combinations of characteristic band selections and inversion model methods identified the optimal approach for modeling the inversion of heavy metal pollution in representative soils characterized by limited sample size and high contaminant concentrations.