Abstract: This study aims to investigate the hyperspectral inversion of soil heavy metal contamination in the Huangshui River Basin. A total of 101 typical soil samples were collected from the study area. By systematically acquiring soil hyperspectral data and measured concentrations of Cr, Zn, and Cu, and integrating spectral preprocessing, characteristic band selection, and machine learning algorithms, quantitative inversion models with high precision were developed to reveal the spatial distribution characteristics of these heavy metals. Five preprocessing methods, including Multiplicative Scatter Correction (MSC), were applied to enhance spectral responses. Sensitive bands were subsequently selected using the Successive Projections Algorithm (SPA), and modeling was performed comparatively using Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Back-Propagation Neural Network (BPNN). The results indicate that the optimal spectral preprocessing methods for Cr, Zn, and Cu were Standard Normal Variate (SNV), Logarithmic Transformation (LR), and Continuum Removal (CR), respectively. SPA identified 10, 12, and 8 characteristic bands for Cr, Zn, and Cu, effectively reducing data dimensionality. The BPNN model achieved the best inversion performance for Cr (R² = 0.943, RMSE = 8.235 mg/kg, RPD = 2.492) and Zn (R² = 0.971, RMSE = 5.370 mg/kg, RPD = 3.714), while the SVR model demonstrated the highest predictive accuracy for Cu (R² = 0.988, RMSE = 0.583 mg/kg, RPD = 8.853). By establishing a systematic workflow encompassing preprocessing, feature selection, and modeling, this study provides methodological support for the rapid monitoring of soil heavy metals in the Huangshui River Basin.