Oil hydrocarbons have a non-negligible impact on crop growth and soil matrix, resulting in crop reduction and even loss of harvest. In order to solve the problem of predicting the concentration of hydrocarbon pollutants in the soil surface, the spectral curve of oil was obtained by fluorescence induction technology, and the wavelet kurtosis was proposed to predict the concentration of polluted oil in the soil surface. Three different oils on the market were compared and analyzed by random forest regression algorithm. The experimental results show that the correlation coefficient R_p and the root mean square deviation RMSD are used to evaluate the prediction results of the three kinds of oil concentrations. The prediction results of gear oil, engine oil and motorcycle oil are increased by 1.2%, 2.2%, 1.9% and 14.9%, 32.4%, 16.8%, respectively. Thirty groups of samples were randomly selected for three kinds of oil, and the recognition accuracy was improved by 6.67%, 6.66% and 9.96%, respectively. At the same time, it is verified that the prediction accuracy of wavelet kurtosis parameter in multiple regression models is improved, and it has high prediction performance. In summary, this study provides a certain reference for the regression model for predicting the concentration of other hydrocarbon pollutants in the soil surface, and provides an effective detection method for agricultural production and the sustainable development of soil environment.