Cobalt (Co) was classified as a group 2B carcinogen by the International Agency for Research on Cancer. It is potentially harmful to the safety of the entire urban ecosystem, and it is particularly important to quickly and accurately detect soil Co content. Hyperspectral techniques have great potential for inversion of soil Co content. 88 surface (0~20 cm) soil samples were collected from Urumqi, Xinjiang, to determine the Co content and original spectral reflectance. The original spectral reflectance was preprocessed and applied with 17 types of transformation, which include the root-mean-square (RMS), the logarithm of the logarithm (LT), the inverse of the logarithm (RL), the inverse of the logarithm (RT), the logarithm of the inverse (AT), the first-order differentiation (FD), the second-order differentiation (SD), the inverse first-order differentiation (RTFD) (RTSD), logarithmic first-order differentiation (LTFD), logarithmic second-order differentiation (LTSD), root-mean-square first-order differentiation (RMSFD), root-mean-square second-order differentiation (RMSSD), logarithmic first-order differentiation of the inverse (ATFD), logarithmic second-order differentiation of the inverse (ATSD), logarithmic first-order differentiation of the inverse (RLFD) and logarithmic second-order differentiation ( RLSD). Then, the Co content and 18 types of soil spectral data were subjected to Pearson correlation analysis (PCC) and CARS to screen the spectral signature variables for modeling. The soil Co content was taken as the dependent variable, and the screened spectral feature variables were taken as independent variables. Based on three algorithms, namely partial least squares regression (PLSR), random forest regression (RFR), and support vector machine regression (SVMR), the hyperspectral inversion models of urban soil Co content were constructed, and the coefficient of determination (R²), the root-mean-square error (RMSE) and the mean absolute error (MAE) were used as the evaluation indexes. Some conclusions can be drawn: The hyperspectral models’ estimation accuracy and stability for urban soil’s Co content are in descending order of the RFR, PLSR, and SVMR models. The best estimation model for Co content is the ATFD-RFR model (R²=0.871，RMSE=0.124，MAE=0.273) which the RPD is 7.90; in this model, compared with the R-RFR model, the R² improved from 0.536 to 0.871, RMSE and MAE reduced by 0.32 and 0.243, respectively. Spectral transform can effectively enhance the spectral features; enhancement of spectral features is most significant with first-order differential transform, among which the RTFD can not only effectively enhance the spectral features of Co but also improve the estimation accuracy of the model very well. The RFR model can be extended in oasis urban soil hyperspectral inversion estimation when the spatial heterogeneity of sample sites is insignificant, and the measured values are low and homogeneous.