Abstract: The NIR spectral detection model of flue-cured tobacco cellulose content under different baking conditions was explored. For 85 samples under different baking conditions, partial least squares regression (PLS) was used to establish the regression model based on all wavelengths. Under common baking, the regression coefficient r=0.9949, RMSE=0.1122 in calibration set; r=0.9234, RMSE=0.4636 in cross validation set; r=0.8982, RMSE=0.6963 in prediction set. Under low temperature baking, r=0.9811, RMSE=0.3279 in calibration set; r=0.9456, RMSE=0.5290 in cross validation set; r=0.9938, RMSE=0.1608 in prediction set. Under high temperature baking, r=0.9128, RMSE=0.4381 in calibration set; r=0.8215, RMSE=0.6162 in cross validation set; r=0.9743, RMSE=0.1986 in prediction set. The results showed that using the partial least squares regression to predict the cellulose content of flue-cured tobacco under different baking conditions was feasible.