Arsenic is one of the heavy metals that seriously harm the human body. The estimation of arsenic content of heavy metals in soil by using hyperspectral technology has great potential for application. However, the applicability and accuracy of the estimation model will vary greatly due to the influence of the region and soil background. The hyperspectral estimation was done in the protection zone of surface water zone of Shijiazhuang. Soil field samples were conducted in the main mining sites and smelting enterprises in the study area and analyzed the heavy metal in the laboratory. After 7-point Savitzky-Golay smoothing of the original spectral reflectance of soil samples, 9 spectral transformations were performed: first derivative (FD), second derivative (SD), reciprocal transformation (RT), reciprocal first derivative (RTFD), reciprocal second derivative (RTSD), absorbance transformation (AT), absorbance transformation and first derivative (ATFD), absorbance transformation and second derivative (ATSD), and continuum removal (CR). Then, the measured content of heavy metal arsenic was correlated with the spectral indexes after spectral transformations, and the maximum sensitive band of each spectral index was extracted. In order to compare the estimation effect of each model, and find the optimal model, the multivariate linear stepwise regression (MLSR), univariate partial least squares regression (U-PLSR) and multivariate partial least squares regression (M-PLSR) method were used to construct the soil heavy metal arsenic content estimation model. The model of arsenic content estimation finally compared the modeling effect through correlation coefficient R², root means square error (RMSE) and statistical value F. The results showed that: most of the samples in the study area were in the critical state of contamination, although some of the soil samples were already mildly contaminated with heavy metal arsenic. The maximum positive and negative correlations existed between CR and ATFD and arsenic content, respectively. Compared with MLSR and U-PLSR, the estimated value of M-PLSR was closest to the measured value, and the fitting samples' R², RMSE and F of M-PLSR were 0.852, 0.147 and 32.384, respectively, which indicated that the integration effect of multivariate transformation modeling was effective. Therefore, the research results can provide a scientific basis for rapid monitoring of heavy metal arsenic pollution in the region.