Abstract: Arsenic (As) is a metalloid element with high carcinogenicity, rendering it particularly important to detect As content in soils in a swift and accurate manner. The study focused on the topsoil in Urumqi City, where 84 soil samples were collected and tested for their As content and original spectral reflectance. This study examined the relationships of As content in the soils with the spectral reflectance under the original spectra and 12 spectral transformations using the Pearson correlation analysis, followed by screening characteristic bands. Hyperspectral models for the inversion of As content in soils were developed using partial least squares regression (PLSR), random forest regression (RFR), and support vector machine regression (SVMR). Finally, the prediction performance of the hyperspectral models was elevated based on the coefficients of determination (R²), root-mean-square errors (RMSEs), and mean absolute errors (MAEs). The results indicated that applying differential transformations to the original spectral data can effectively enhance the spectral features and improve the correlation between spectral reflectance and As content in soils. The prediction performance of the hyperspectral models decreased in the order of RFR, SVMR, and PLSR. The RFR model based on root-mean-square second order differentiation (RMSSD-RFR) exhibited the best fitting effects and the highest prediction stability, with R² of 0.821, a RMSE of 0.143 mg/kg, and a MAE of 0.523 mg/kg. This study provides a scientific basis for developing hyperspectral models for the inversion of As content in soils in an oasis city.