Fourier transform infrared spectroscopy is an important method for monitoring air pollution emissions from pollution sources. Automatic baseline correction method for gas spectra is of great significance to air pollution monitoring applications, such as rapid detection and long-term on-line monitoring. One difficulty in the current automatic baseline correction is accurately correcting the spectra, which include broad peaks. The broad peakscontain low-frequency content in the frequency domain; thus, the method for extracting baseline information based on low-frequency filtering is prone to baseline distortion because of the difficulty in selecting the appropriate separation parameter. Automatically identifying the baseline point and fitting the baseline of the spectrum based on a preset baseline function can prevent the selection of separation conditions; however, the result of baseline correction is highly sensitive to the baseline function adopted. If the degree of freedom in the baseline function is excessively small, the baseline function cannot fit the baseline drift in the spectra accurately, and the error will be considerable after baseline correction. Meanwhile, if the degree of freedom in the baseline function is excessively large, in particular, when a false degree of freedom does not exist in the natural baseline drift, the fitted baseline may have baseline distortion. Many types of baseline functions exist, including linear, polynomial, spline interpolation, and exponential functions. At present, consensus is lacking regarding the selection criteria for baseline functions. In this study, we proposed a baseline function for gas spectra for extractive atmospheric monitoring based on the degree of freedom of the natural baseline drift; we aimed to avoid false degrees of freedom or lack of necessary degrees of freedom in the baseline function. We found that the degrees of freedom of major baseline drift in the gas spectrum can be approximated in specific order terms of wavenumbers (0, 1st-, 2nd-, and 4th-order terms). An automatic baseline correction method based on a polynomial baseline function with above (0, 1st-, 2nd-, and 4th-) order terms was proposed in this study. In the experiment, a measured air spectrum, which contained broad peaks of water vapor, was used as a sample to test the performance of the baseline correction method. The baseline correction result of the proposed automatic baseline correction method was compared with the that of two types of iterative polynomial fitting methods that were proposed by Lieber and Mahadeven-Jansen (LMJ) and by Liu and Koenig (LK). The experiment results indicated that compared with the LMJ and LK methods, the proposed method avoided the baseline distortion in the best possible manner, and the proposed method also showed the lowest average variance between the corrected baseline and the absorbance zero line. Our research showed that in automatic baseline correction, an effective baseline correction result can be obtained by establishing the baseline function with the freedom of the natural baseline drift.