Hyperspectral remote sensing technology is an effective method to monitor petroleum contamination. It is mainly used in offshore oil spill, while few study has been focusing on soil petroleum-hydrocarbon contamination. In this case, due to the shortage of research on soil petroleum-hydrocarbon contamination, three kinds of petroleum hydrocarbons are selected, including diesel, gasoline and motor oil, to characterize the absorption features of petroleum hydrocarbons in spectra yielded from contaminated purplish soils in condition of different types and different concentrations of petroleum hydrocarbons and extract the spectral absorption characteristic band of soil contaminated by petroleum hydrocarbon. Based on this, seven kinds of spectral transformations and correlation analysis were conducted to select the most sensitive spectral variables with petroleum hydrocarbon content. The estimation model was established via univariate regression and multiple stepwise linear regression (SMLR) respectively and verified ultimately. The results show that the spectral signatures of soil contaminated by diesel, motor oil and gasoline are in the vicinity of 1 200, 1 700 and 2 300 nm, and the absorption depth is shown as follows: motor oil>gasoline>diesel. Multiple stepwise linear regression method is superior to univariate regression method. The coefficients of determination (R2) of diesel, motor oil and gasoline are greater than 0.95. Moreover, the root mean square error of calibration (RMSEC) is less than 0.47 and the root mean square error of validation (RMSEC) is less than 0.56, which demonstrates higher estimation accuracy.