The petroleum pollutant is an important factor causing air pollution problems such as haze. The de-noising effectiveness is the focus in petroleum pollutant detection by fluorescence spectroscopy. A fluorescence spectrum de-noising method for low concentration petroleum pollutants combining the empirical model decomposition (EMD) and the lifting wavelet transform (LWT) is proposed. The EMD method can filter the noise in weak fluorescence signal adaptively, but the first intrinsic mode function (IMF) contains a too wide frequency range, and thus the de-noising accuracy and effectiveness is reduced. LWT is introduced to realize more precise decomposition of IMF1, extract more useful information from IMF1, and improve separation effect of signal and noise. The three de-noising methods, EMD-LWT, EMD and LWT, are applied to kerosene fluorescence spectrum detection, respectively. The simulation results show that the EMD-LWT method makes the signal-to-noise ratio，root mean square error significantly improved compared with only EMD or LWT used, verifying the effectiveness and feasibility of the proposed method.