Fourier transform infrared spectroscopy (FTIR) can rapidly and nondestructively measure the composition and content of organic matter. As an important organic chemical raw material, ethylene is widely used in the manufacturing of bulk chemicals such as plastics, alcohols and fibers. However, due to its volatility, ethylene is harmful to the environment and human body. To improve the accuracy of FTIR in the detection model of ethylene concentration, an improved IRIV-SA infrared spectral wavenumber optimization algorithm is proposed based on the advantages of iteratively retains informative variables method (IRIV) and simulated annealing algorithm(SA). On the basis of stable selection of a large number of spectral characteristic wavenumbers by IRIV algorithm, this method uses SA to further screen a small number of effective characteristic wavenumbers, so as to reduce the complexity of the model and improve the detection accuracy of organic matter spectrum. In the experiment, IRIV-SA is used to select the wavenumber of the concentration of ethylene infrared spectrum at first, and the number of characteristic wavenumbers obtained is reduced from 271 to 5, then the characteristic wavenumber is used for modeling. The results show that the correlation coefficient and root mean square error of validation set are 0.9989 and 0.3943, and the correlation coefficient and root mean square error of prediction set are 0.9978 and 0.6652, which indicates that the modeling accuracy of the proposed algorithm is significantly improved compared with that of the whole spectrum modeling. To further verify the effectiveness of the improved algorithm, IRIV, SA, CARS (competitive adaptive reweighted sampling algorithm), SPA (successive projections algorithm), IRIV-CARS and IRIV-SPA wavenumber selection models are established for comparative experiments on the same data set. The comparison results show that IRIV-SA algorithm is superior to the above six wavenumber selection methods, and is an effective feature wavenumber selection method.