In order to improve the prediction accuracy and model stability for quantitative calibration in near infrared spectroscopy, the random frog wavelength selection method combined with partial least squares (PLS) was employed to establish the prediction model of near infrared spectra for petroleum ether extracts in flue-cured tobacco leaves from various geographical regions. A combination of the standard normal variate and the first derivative was selected as the spectral pre-treatment method. The samples were subdivided into a calibration set (285 samples) and a validation set (95 samples) according to 3:1 by the Kennard-Stone (KS) algorithm. The random frog method was used for spectral variable selection. The partial least squares regression (PLSR) calibration models were established to predict the content of petroleum ether extracts, and compared with the results of full spectra. The results showed that better prediction was obtained by the random frog method compared to full spectra PLS modelling, moving window PLSR (MW-PLSR) and Monte-Carlo uninformative variable elimination PLSR (MC-UVEPLSR). The determination coefficient of calibration set (R²), root mean square error of cross validation (RMSECV), determination coefficient of validation set (r²) and root mean square error of prediction (RMSEP) were used to evaluate the quality of the model. R², RMSECV, r² and RMSEP of the full spectrum PLS model were 0.9902, 0.0896 %, 0.8573, and 0.3415 %, respectively. While R², RMSECV, r² and RMSEP of the random frog PLS model were 0.9934, 0.0830 %, 0.9700 and 0.1547 %, respectively. The result showed that using a variable selection method such as the random frog method could effectively select the characteristic wavelengths of NIR spectrum to improve the model applicability and robustness. The proposed method could effectively simplify the model and was suitable for fast and reliable determination of petroleum ether extracts in flue-cured tobacco leaves.